Time estimates for one's own tasks: errors, kinds of task, and learning an estimate-to-actual ratio¶

Research for Sioul's estimate-to-actual learning (started 2026-10-05). Peer-reviewed sources first, then practice and tools. Each source: reference, finding, study type and size, strength. References checked against Crossref/PubMed/publisher unless marked "unverified".

Strength scale used below: strong (meta-analysis, large review, or replicated experiments), moderate (several experiments, one lab, or one large field dataset), weak (one small study, case study), opinion (practitioner writing, books without data).

Status: complete 2026-10-05 (see end).

1. The planning fallacy and its mechanisms¶

1.1 Buehler, Griffin & Ross 1994 — the founding studies¶

Buehler R, Griffin D, Ross M (1994). Exploring the "planning fallacy": Why people underestimate their task completion times. Journal of Personality and Social Psychology 67(3):366–381. doi:10.1037/0022-3514.67.3.366 (Crossref-verified; full text read via an MIT course copy).

• Found (Study 1, 37 honours-thesis students, means on 33): predicted 33.9 days, actual 55.5 days (ratio about 1.6); 29.7 % finished by their own predicted date; even the "as poorly as it possibly could" prediction (48.6 days) was met by only 48.7 %. Yet predicted and actual days correlated r = .77: "Predictions can be informative even in the presence of a marked prediction bias."
• Studies 2–5 (academic and everyday tasks): people's thoughts focused on future plans, not on past experience. Study 4: "the optimistic bias was eliminated for Ss instructed to connect relevant past experiences with their predictions" (recalling alone was not enough). Study 5: observers overestimated others' completion times and used past experience more.
• Strength: strong as a phenomenon (replicated many times), but it concerns completion dates (calendar), not time on task. See 1.6.
• For Sioul: the person's estimate carries real information (r = .77) even when biased — a correction of the estimate, rather than a replacement for it, is the right shape.

1.2 Kahneman & Tversky 1979; Kahneman & Lovallo 1993 — inside versus outside view¶

• Kahneman D, Tversky A (1979). Intuitive prediction: biases and corrective procedures. TIMS Studies in Management Science 12:313–327. (Book-series chapter, not in Crossref; widely cited, e.g. by Buehler 1994 and Flyvbjerg 2006. Treated as verified by citation, not seen.) Coins "planning fallacy" and proposes a five-step correction: pick a reference class, get its distribution, make the intuitive estimate, judge how predictable the case is, then move the intuitive estimate toward the class average in proportion to the unpredictability — a shrinkage rule.
• Kahneman D, Lovallo D (1993). Timid choices and bold forecasts: a cognitive perspective on risk taking. Management Science 39(1):17–31. doi:10.1287/mnsc.39.1.17 (verified). Names the "inside view" (plan the case from its own details) versus the "outside view" (look at how similar cases turned out); forecasts from the inside view are too bold.
• Strength: opinion/theory from very credible authors; the corrective procedure itself was not tested in these papers. It is the root of reference-class forecasting (section 4).

1.3 Roy, Christenfeld & McKenzie 2005 and follow-ups — biased memory of past durations¶

• Roy MM, Christenfeld NJS, McKenzie CRM (2005). Underestimating the duration of future events: memory incorrectly used or memory bias? Psychological Bulletin 131(5):738–756. doi:10.1037/0033-2909.131.5.738 (verified, PMID 16187856). Review. "People appear to underestimate future event duration because they underestimate past event duration." Not exclusive with the optimism account.
• Roy MM, Christenfeld NJS (2007). Bias in memory predicts bias in estimation of future task duration. Memory & Cognition 35(3):557–564. doi:10.3758/BF03193294 (verified, PMID 17691153). Two experiments: underestimation of future duration disappears when the task is novel; past and future estimates show the same bias; experience with the task and delay before estimating shift both in the same way.
• Roy MM, Christenfeld NJS (2008). Effect of task length on remembered and predicted duration. Psychonomic Bulletin & Review 15(1):202–207. doi:10.3758/PBR.15.1.202 (verified, PMID 18605504). Tasks of 1–15 min: "task duration tended to be overestimated for short tasks (less than 2 min) and underestimated for long tasks for both remembered and predicted duration" (Vierordt's law extended to predictions).
• Roy MM, Mitten ST, Christenfeld NJS (2008). Correcting memory improves accuracy of predicted task duration. Journal of Experimental Psychology: Applied 14(3):266–275. doi:10.1037/1076-898X.14.3.266 (verified, PMID 18808280). Three experiments (paper counting, essay writing, field sample at restaurants/arcades): giving people the measured duration of their own previous similar task, or the average duration for others, before they predict "increased predictive accuracy"; "people do well when they rely not on memory of past task duration but instead on measures of actual duration".
• Strength: moderate — one lab, several small experiments (sizes not checked), short lab tasks (minutes). Directly relevant to an app that measures durations: the app can replace memory with measurement.

1.4 Unpacking and segmentation¶

• Kruger J, Evans M (2004). If you don't want to be late, enumerate: unpacking reduces the planning fallacy. Journal of Experimental Social Psychology 40(5):586–598. doi:10.1016/j.jesp.2003.11.001 (verified; full text paywalled, details from secondary summaries). Five experiments (getting ready for a date, holiday shopping, formatting a document, cooking): listing the steps raised predicted times and made them more accurate; the more components a task had, the more unpacking helped, and it did little for tasks with few components (support theory: unpacked descriptions get more weight). Moderate.
• Forsyth DK, Burt CDB (2008). Allocating time to future tasks: the effect of task segmentation on planning fallacy bias. Memory & Cognition 36(4):791–798. doi:10.3758/MC.36.4.791 (verified, PMID 18604961). Three experiments: time allocated to a whole task was "significantly smaller than the summed time allocated to the individual subtasks" (segmentation effect); not explained by rounding. Moderate.
• Hadjichristidis C, Summers B, Thomas KE (2014). Unpacking estimates of task duration: the role of typicality and temporality. Journal of Experimental Social Psychology 51:45–50. doi:10.1016/j.jesp.2013.10.009 (verified). Shows boundary conditions of the unpacking effect (details to be checked; abstract paywalled — finding not confirmed here).
• Consequence for an app: the ratio depends on the granularity at which the person estimates. Estimates of subtasks sum to more than the estimate of the whole; a ratio learned on whole tasks does not transfer to subtasks.

1.5 Scenario thinking and motivation¶

• Newby-Clark IR, Ross M, Buehler R, Koehler DJ, Griffin D (2000). People focus on optimistic scenarios and disregard pessimistic scenarios while predicting task completion times. Journal of Experimental Psychology: Applied 6(3):171–182. doi:10.1037/1076-898X.6.3.171 (verified, PMID 11014050). Five experiments: writing a pessimistic scenario "did not affect predictions or accuracy"; people did use pessimistic scenarios when predicting someone else's times. "Pessimistic-scenario generation may not be an effective debiasing technique for personal predictions." Moderate. Implication: asking the person to imagine what could go wrong is not a fix; using measured history is.
• Buehler R, Griffin D, MacDonald H (1997). The role of motivated reasoning in optimistic time predictions. Personality and Social Psychology Bulletin 23(3):238–247. doi:10.1177/0146167297233003 (verified). Two studies: people expecting a tax refund were more over-optimistic about when they would finish their tax forms; in the lab, money for early completion made predictions more optimistic, drew attention to detailed plans and away from relevant past experience. Wanting a task done soon biases the estimate. Moderate.

1.5b Optimism, implementation plans, Hofstadter¶

• Weinstein ND (1980). Unrealistic optimism about future life events. JPSP 39(5):806–820. doi:10.1037/0022-3514.39.5.806 (verified). People rate their own chances of good events above, and of bad events below, their peers'. Sharot T (2011). The optimism bias. Current Biology 21(23):R941–R945. doi:10.1016/j.cub.2011.10.030 (verified), short review. Strong as a general bias; only one of several causes of time misprediction (1.6).
• Koole S, van't Spijker M (2000). Overcoming the planning fallacy through willpower: effects of implementation intentions on actual and predicted task-completion times. European Journal of Social Psychology 30(6):873–888. doi:10.1002/1099-0992(200011/12)30:6<873::AID-EJSP22>3.0.CO;2-U (verified). Planning where and when to do a task made predictions more optimistic but completion improved even more, so the bias shrank; the gain was "mediated by a smaller number of interruptions". Moderate (one experiment). Interruptions are part of what the ratio measures.
• Hofstadter DR (1979). Gödel, Escher, Bach. Basic Books. "Hofstadter's Law: It always takes longer than you expect, even when you take into account Hofstadter's Law." Aphorism, not evidence — and the data above show it is not always true for small tasks.

1.6 Size of the error, and when it is optimistic at all¶

• Halkjelsvik T, Jørgensen M (2012). From origami to software development: a review of studies on judgment-based predictions of performance time. Psychological Bulletin 138(2):238–271. doi:10.1037/a0025996 (verified, PMID 22061688). The key review for Sioul because it is about time on task (performance time), not deadlines. Quotes from the abstract: "underestimation was more frequently reported than overestimation in studies from the engineering and management literature. However, this was not the case in studies from the psychology literature." "We also question the recurrent finding that small tasks are overestimated and large tasks are underestimated, as this to some extent can be a statistical artifact caused by random error." Also lists anchoring, incentives, decomposition, request formats, and the need to separate performance-time from completion-time predictions. Strong (large systematic review across psychology, engineering and management).
• Halkjelsvik T, Jørgensen M (2018). Time Predictions: Understanding and Avoiding Unrealism in Project Planning and Everyday Life. Simula SpringerBriefs on Computing 5, Springer, open access (CC BY 4.0). doi:10.1007/978-3-319-74953-2 (verified; read in full text). Points used here:
  • Typical average error (unsigned) across domains: "around 20–30%". Across all reported tasks, "we did not find a general tendency towards either overoptimistic or overpessimistic time predictions"; lab studies of everyday tasks were on average unbiased; small software tasks were unbiased or overestimated; projects of 100+ work hours had a median overrun of about 20 %.
  • Selection effect: we start tasks more readily when we think they are quick, and only completed tasks get scored, so optimism is over-represented in what we remember.
  • Time-usage distributions are right-skewed ("nearly all"); extreme deviations are almost always overruns; "your efficiency is rarely as extreme as your inefficiency." Example: 443 IT tasks, more tasks overestimated (215) than underestimated (196), yet the total ran 15 % over (2723 h predicted, 3130 h actual).
  • "For the proper addition of time predictions, you should add the predicted mean value of each subtask", not the most likely or median values; sums of skewed times are larger than sums of typical times. Monte Carlo without modelling dependencies gives falsely narrow totals.
  • People's own min/max intervals are far too narrow: 98 % intervals contained the actual time in 57 % of cases (Connolly & Dean 1997).
  • Predictions change the work (Parkinson): too-high predictions can lower productivity; "Rewarding accurate time predictions is usually not a good idea" because people stretch or shrink work to match.
  • Who predicts well: "previous time prediction overoptimism was the best indicator of future time prediction overoptimism" (Jørgensen, Faugli & Gruschke 2007, software developers), but with a large random element. Highly relevant experience helps; general experience little. Training by feedback and reflection (PSP; lessons-learned sessions) did not improve accuracy in independent studies (Abrahamsson & Kautz 2002; Prechelt & Unger 2001; Jørgensen & Gruschke 2009).
  • Experience can increase optimism while improving accuracy: pianists predicted +48 % for newly learned pieces, +11 % for well learned, −19 % for extremely well learned (Boltz, Kupperman & Dunne 1998); origami novices overestimated, practiced folders underestimated (Roy & Christenfeld 2007).
  • Strength: strong as a synthesis; the authors flag where evidence rests on one study.

2. Time perception and estimation in ADHD, depression, anxiety, fatigue¶

2.1 ADHD: what the lab timing literature shows¶

• Barkley RA (1997). Behavioral inhibition, sustained attention, and executive functions: constructing a unifying theory of ADHD. Psychological Bulletin 121(1):65–94. doi:10.1037/0033-2909.121.1.65 (verified, PMID 9000892). Theory: ADHD as a deficit of inhibition that impairs working memory and self-regulation, including the sense of time. "Time blindness" is Barkley's clinical shorthand from his books and talks, not a measured construct. Theory, highly cited.
• Barkley RA, Murphy KR, Bush T (2001). Time perception and reproduction in young adults with attention deficit hyperactivity disorder. Neuropsychology 15(3):351–360. doi:10.1037/0894-4105.15.3.351 (verified, PMID 11499990). 104 adults with ADHD vs 64 controls, intervals of seconds. ADHD group gave larger verbal estimates (difference vanished after controlling for IQ), made shorter reproductions and larger reproduction errors; anxiety added errors at 12 s. Moderate.
• Toplak ME, Dockstader C, Tannock R (2006). Temporal information processing in ADHD: findings to date and new methods. Journal of Neuroscience Methods 151(1):15–29. doi:10.1016/j.jneumeth.2005.09.018 (verified, PMID 16378641). Review: evidence for problems in "duration discrimination, duration reproduction, and finger tapping"; methods too varied to integrate; "further replication is needed". Narrative review.
• Noreika V, Falter CM, Rubia K (2013). Timing deficits in ADHD: evidence from neurocognitive and neuroimaging studies. Neuropsychologia 51(2):235–266. doi:10.1016/j.neuropsychologia.2012.09.036 (verified, PMID 23022430). Review: consistent impairment in motor timing, perceptual timing and "temporal foresight"; most consistent in sensorimotor synchronisation, duration discrimination, reproduction and delay discounting; deficits survive control for working memory and attention. Narrative review, strong for seconds-scale lab tasks.
• Marx I, Cortese S, Koelch MG, Hacker T (2022). Meta-analysis: altered perceptual timing abilities in ADHD. J Am Acad Child Adolesc Psychiatry 61(7):866–880. doi:10.1016/j.jaac.2021.12.004 (verified, PMID 34923055). 55 studies, milliseconds to several seconds: worse discrimination, "more variability in estimating the duration of stimuli lasting several seconds", estimation/production errors "indicative of an accelerated internal clock". Strong (meta-analysis), but seconds only.
• Zheng Q, Wang X, Chiu KY, Shum KK (2022). Time perception deficits in children and adolescents with ADHD: a meta-analysis. Journal of Attention Disorders 26(2):267–281. doi:10.1177/1087054720978557 (verified, PMID 33302769). 27 studies, 1620 ADHD vs 1249 controls: less accurate (g > 0.40), less precise (g = 0.66), "higher tendency to overestimate time". Strong, children only.
• Metcalfe KB, McFeaters CD, Voyer D (2024). Time-perception deficits in ADHD: a systematic review and meta-analysis. Developmental Neuropsychology 49(1):1–24. doi:10.1080/87565641.2023.2293712 (verified, PMID 38145491). 824 effect sizes across the lifespan, mean g = 0.688; moderated by age and working memory. Strong.
• Nejati V, Yazdani S (2020). Time perception in children with ADHD: does task matter? A meta-analysis. Child Neuropsychology 26(7):900–916. doi:10.1080/09297049.2020.1712347 (verified, PMID 32757699). 12 studies; problems clearer for prospective tasks and longer intervals. Moderate (small meta-analysis).
• Mette C (2023). Time perception in adult ADHD: findings from a decade — a review. IJERPH 20(4):3098. doi:10.3390/ijerph20043098 (verified, PMID 36833791). Studies in adults are "very scarce"; some show deficits in estimation, reproduction and time management, others "were unable to demonstrate a clear association". Narrative review; shows the adult evidence is thin.
• Ptáček R, Vňuková M, Děchtěrenko F, et al. (2022). ADHD and time perception in adults: do adults with different ADHD symptomatology severity perceive time differently? Medical Science Monitor 28:e936849. doi:10.12659/MSM.936849 (verified, PMID 35841140). Groups drawn from a national sample of 1518: no meaningful difference by symptom severity in duration estimation or discrimination (interaction effect size ηp² < 0.001), although people with severe symptoms report more time-management trouble in daily life. Moderate; a useful null.
• Prevatt F, Proctor B, Baker L, Garrett L, Yelland S (2011). Time estimation abilities of college students with ADHD. Journal of Attention Disorders 15(7):531–538. doi:10.1177/1087054710370673 (verified, PMID 20679155). 20 vs 20 students, one novel complex task resembling study work: the ADHD group differed on completion time, prospective estimate, retrospective estimate and both errors (controlling for ability); per Walg et al. 2017 the ADHD group overestimated durations prospectively and retrospectively. Weak (n = 40, one task) — but it is the closest study to "predict how long this task will take".
• Walg M, Hapfelmeier G, El-Wahsch D, Prior H (2017). The faster internal clock in ADHD is related to lower processing speed. European Child & Adolescent Psychiatry 26(10):1177–1186. doi:10.1007/s00787-017-0971-5 (verified, PMID 28283836). 50 boys with ADHD vs 53 clinical controls, retrospective estimates of several minutes: ADHD estimates "longer and less accurate". Moderate.
• Kofler MJ, Rapport MD, Sarver DE, et al. (2013). Reaction time variability in ADHD: a meta-analytic review of 319 studies. Clinical Psychology Review 33(6):795–811. doi:10.1016/j.cpr.2013.06.001 (verified, PMID 23872284). Greater trial-to-trial variability (children g = 0.76, adults g = 0.46), not slower mean speed once variability is accounted for; adults with ADHD were indistinguishable from clinical control groups. Strong, but milliseconds. Analogy only: expect wider spread of actual/estimate ratios, not necessarily a shifted mean.
• Hupfeld KE, Abagis TR, Shah P (2019). Living "in the zone": hyperfocus in adult ADHD. ADHD Attention Deficit and Hyperactivity Disorders 11(2):191–208. doi:10.1007/s12402-018-0272-y (verified, PMID 30267329). Two samples (251; 372, preregistered): higher ADHD symptoms go with more frequent hyperfocus in school, hobbies and screen time. Moderate (self-report). Relevant: absorbing tasks overrun because stopping fails, not because the estimate was wrong.
• Niermann HC, Scheres A (2014). The relation between procrastination and symptoms of ADHD in undergraduate students. Int J Methods Psychiatr Res 23(4):411–421. doi:10.1002/mpr.1440 (verified, PMID 24992694). n = 54: only inattention correlated with general procrastination. Weak.
• What is missing: no peer-reviewed study found that measures the planning fallacy (predicted vs actual time for real tasks lasting minutes to hours) in adults with ADHD. Searched PubMed, OpenAlex and the web with "ADHD planning fallacy", "ADHD task duration prediction", "time management". Popular sources (ADDitude, blogs) assert that people with ADHD underestimate more, sometimes with numbers ("50 % more") that trace to no study — opinion.
• Reading for Sioul: the lab evidence is about seconds; it says more variable and, if anything, longer felt durations (faster internal clock), not systematically shorter predictions. Adult evidence is thin and partly null. So the app should not assume a bigger multiplier for ADHD; it should expect a wider spread and learn the person's own ratio. Self-reported time trouble is real (Ptáček 2022) even where lab tasks show none, which points to planning, starting, stopping and interruptions rather than perception.

2.2 Depression, anxiety, burnout, fatigue¶

• Thönes S, Oberfeld D (2015). Time perception in depression: a meta-analysis. Journal of Affective Disorders 175:359–372. doi:10.1016/j.jad.2014.12.057 (verified, PMID 25665496). 16 studies, 433 patients, 485 controls: patients feel time "going by less quickly" (g = 0.66), but "no significant effects of depression in the four time perception tasks" (estimation, production, reproduction, discrimination); a tendency to overproduce short and underproduce long durations. Strong for "feels slow, measures normal".
• Bschor T, Ising M, Bauer M, et al. (2004). Time experience and time judgment in major depression, mania and healthy subjects: a controlled study of 93 subjects. Acta Psychiatrica Scandinavica 109(3):222–229. doi:10.1046/j.0001-690X.2003.00244.x (verified, PMID 14984395). Depressed patients felt time slow, manic patients fast, yet both overestimated longer spans in judgment tasks. Moderate.
• Gil S, Droit-Volet S (2009). Time perception, depression and sadness. Behavioural Processes 80(2):169–176. doi:10.1016/j.beproc.2008.11.012 (verified, PMID 19073237). Higher depression scores, shorter judged durations (sub-second to 1.6 s), consistent with a slower internal clock. Weak–moderate (one study, milliseconds).
• Moore MT, Fresco DM (2012). Depressive realism: a meta-analytic review. Clinical Psychology Review 32(6):496–509. doi:10.1016/j.cpr.2012.05.004 (verified, PMID 22717337). 75 studies, 7305 people: "a small overall depressive realism effect (d = −.07)"; both depressed (d = .14) and non-depressed (d = .29) people show positive bias. Strong. Implication: do not expect depressed or exhausted users to be "realistic" planners by default; optimism is only slightly smaller.
• Bar-Haim Y, Kerem A, Lamy D, Zakay D (2010). When time slows down: the influence of threat on time perception in anxiety. Cognition & Emotion 24(2):255–263. doi:10.1080/02699930903387603 (verified). 29 anxious vs 29 non-anxious: anxious people felt 2-s threat images last longer; the difference vanished at 4–8 s. Weak (small, seconds).
• Sarigiannidis I, Grillon C, Ernst M, Roiser JP, Robinson OJ (2020). Anxiety makes time pass quicker while fear has no effect. Cognition 197:104116. doi:10.1016/j.cognition.2019.104116 (verified, PMID 31883966). Three experiments with threat of shock: induced anxiety made people underestimate durations (d = 0.68); fear effects did not replicate. Moderate. With Bar-Haim, shows that the direction of emotion effects on seconds-scale timing is not settled.
• Deligkaris P, Panagopoulou E, Montgomery AJ, Masoura E (2014). Job burnout and cognitive functioning: a systematic review. Work & Stress 28(2):107–123. doi:10.1080/02678373.2014.909545 (verified). Burnout linked to poorer executive function, attention and memory in several studies, but findings are heterogeneous and samples small (as later summarised by Koutsimani & Montgomery 2022, Front Psychiatry 13:775606, doi:10.3389/fpsyt.2022.775606, verified). Moderate–weak.
• Goudsmit EM, Nijs J, Jason LA, Wallman KE (2012). Pacing as a strategy to improve energy management in myalgic encephalomyelitis/chronic fatigue syndrome: a consensus document. Disability and Rehabilitation 34(13):1140–1147. doi:10.3109/09638288.2011.635746 (verified, PMID 22181560). Pacing — keeping activity within current limits, adjusted to the person — is the strategy patients rate most helpful and may prevent post-exertional crashes. Consensus/review; evidence modest. Jason LA, Brown M, Brown A, et al. (2013). Energy conservation/envelope theory interventions to help patients with ME/CFS. Fatigue: Biomedicine, Health & Behavior 1(1–2):27–42. doi:10.1080/21641846.2012.733602 (verified, PMID 23504301): staying inside one's "energy envelope" improved quality of life in the authors' studies. Moderate–weak (one group). For a planner: the day's capacity, not only each task's duration, must be conservative; overruns cost more than for healthy users.
• What is missing: no study found on time perception or task-duration prediction in burnout, ME/CFS or long COVID (PubMed searches: burnout/"exhaustion disorder" × "time perception"/"time estimation"; "chronic fatigue syndrome"/"mental fatigue"/"sleep deprivation" × "time perception" returned only seconds-scale sports or sleep-loss studies). Anything an app assumes here is extrapolation.
• Reading for Sioul: mood and anxiety change how time feels far more than how it is judged; the measurable effect on duration judgments is small and inconsistent. Exhaustion is more likely to change actual durations (slower work, more breaks, worse days) than the estimate. So the ratio should be allowed to move with the person's state (short window, day effects), and the app should not infer a condition from the ratio.

2.3 How duration is judged — background that matters for the data¶

• Block RA, Zakay D (1997). Prospective and retrospective duration judgments: a meta-analytic review. Psychonomic Bulletin & Review 4(2):184–197. doi:10.3758/BF03209393 (verified, PMID 21331825). 20 experiments: judgments made while watching the time (prospective) are longer and less variable than judgments made afterwards from memory (retrospective); different processes (attention versus memory). Strong.
• Block RA, Hancock PA, Zakay D (2010). How cognitive load affects duration judgments: a meta-analytic review. Acta Psychologica 134(3):330–343. doi:10.1016/j.actpsy.2010.03.006 (verified, PMID 20403583). 117 experiments: with higher cognitive load the prospective ratio (felt/real) decreases — absorbing work feels shorter — while the retrospective ratio increases. Strong. Absorbing or demanding tasks are where felt time and clock time part most.
• Jazayeri M, Shadlen MN (2010). Temporal context calibrates interval timing. Nature Neuroscience 13(8):1020–1026. doi:10.1038/nn.2590 (verified, PMID 20581842). Reproduced intervals regress toward the mean of the intervals recently experienced; a Bayesian observer fits. Strong (lab, sub-second to ~1 s). The same central-tendency pull is the likely root of "short tasks overestimated, long ones underestimated" (Vierordt), and means the person's estimates will drift with what they have lately been doing.
• Consequence: manually entered "actual" times are retrospective judgments and carry the memory bias Roy et al. describe; timer-measured times do not. Weight them differently or exclude them.

3. Task affect, difficulty and familiarity¶

Short answer: the direction of estimation error by enjoyment is not established for real tasks; what is established is that affect changes felt time, that aversive tasks get delayed and fragmented, that absorbing tasks are hard to stop, and that length and familiarity change the error.

3.1 Enjoyable versus aversive tasks¶

• Droit-Volet S, Meck WH (2007). How emotions colour our perception of time. Trends in Cognitive Sciences 11(12):504–513. doi:10.1016/j.tics.2007.09.008 (verified, PMID 18023604). Review: "time seems to fly when we are having fun and drags when we are bored"; arousal and valence act on attention and clock speed in both directions. Review, lab scale.
• Gable PA, Poole BD (2012). Time flies when you're having approach-motivated fun. Psychological Science 23(8):879–886. doi:10.1177/0956797611435817 (verified, PMID 22760888). Three experiments: positive states with high approach motivation (wanting) shorten perceived time; arousal alone does not. Moderate. "Shortening of time perception … may prolong approach-motivated behavior" — i.e. wanted tasks run on.
• Sackett AM, Meyvis T, Nelson LD, Converse BA, Sackett AL (2010). You're having fun when time flies: the hedonic consequences of subjective time progression. Psychological Science 21(1):111–117. doi:10.1177/0956797609354832 (verified, PMID 20424031). Seven studies: when people believe time passed quickly they rate the task as more engaging. The causal arrow runs both ways. Moderate.
• Danckert JA, Allman AA (2005). Time flies when you're having fun: temporal estimation and the experience of boredom. Brain and Cognition 59(3):236–245. doi:10.1016/j.bandc.2005.07.002 (verified, PMID 16168546). Boredom-prone people made larger timing errors; low-boredom people tended to underestimate. Weak (small, lab).
• Zakay D (2014). Psychological time as information: the case of boredom. Frontiers in Psychology 5:917. doi:10.3389/fpsyg.2014.00917 (verified, PMID 25191296). Theory: boredom slows felt time as a signal to recruit effort. Opinion/theory.
• Tonietto GN, VanEpps EM, Malkoc SA, Maglio SJ (2022). Time will fly during future fun (but drag until then). Journal of Consumer Psychology 32(3):475–483. doi:10.1002/jcpy.1247 (verified). Four studies: people judge future positive events of equal objective length as shorter than negative ones — a prospective judgment, which is what an estimate is. Moderate (scheduled events, not effortful tasks).
• Pezzo MV, Litman JA, Pezzo SP (2006), cited in 4.1: present-focused, carefree people predicted early completion of boring tasks and finished later. Weak–moderate.
• Buehler, Griffin & MacDonald 1997 (1.5): wanting a task finished quickly makes the prediction more optimistic and pulls attention away from past experience. Brunnermeier MK, Papakonstantinou F, Parker JA (2017). Optimal time-inconsistent beliefs: misplanning, procrastination, and commitment. Management Science 63(5):1318–1340 (online 2016). doi:10.1287/mnsc.2015.2360 (verified). Economic theory: optimism about a task's ease is "chosen" because anticipating an easy task feels better; incentives for speed worsen it, incentives for accuracy reduce it. Its premise "faced with an unpleasant task, people tend both to underestimate the time … and to postpone" is asserted from the planning-fallacy literature, not tested by pleasant/unpleasant contrast. Theory.
• Steel P (2007). The nature of procrastination: a meta-analytic and theoretical review of quintessential self-regulatory failure. Psychological Bulletin 133(1):65–94. doi:10.1037/0033-2909.133.1.65 (verified, PMID 17201571). 691 correlations: "strong and consistent predictors of procrastination were task aversiveness, task delay, self-efficacy, and impulsiveness" plus conscientiousness facets. Strong.
• Sirois F, Pychyl T (2013). Procrastination and the priority of short-term mood regulation: consequences for future self. Social and Personality Psychology Compass 7(2):115–127. doi:10.1111/spc3.12011 (verified). Procrastination as short-term mood repair when facing aversive tasks. Review/theory.
• Blunt AK, Pychyl TA (2000). Task aversiveness and procrastination: a multi-dimensional approach to task aversiveness across stages of personal projects. Personality and Individual Differences 28(1):153–167. doi:10.1016/S0191-8869(99)00091-4 (verified; abstract not read, so only the title's claim is used: aversiveness measured across the stages of personal projects relates to procrastination). Not rated.
• Pychyl TA, Morin RW, Salmon BR (2000). Procrastination and the planning fallacy: an examination of the study habits of university students. Journal of Social Behavior and Personality 15(5):135–150 (special issue; not in Crossref; confirmed via library catalogue record). 32 students' study plans for the 8 days before two midterms versus their logs: "overall students did not demonstrate optimistic biases". Weak (n = 32), a useful null: the plan-versus-log gap in a real setting was not one-directional.
• Breig Z, Gibson M, Shrader JG (2020). Why do we procrastinate? Present bias and optimism. IZA Discussion Paper 13060; SSRN doi:10.2139/ssrn.3445511 (working paper, not peer-reviewed). Real-effort field experiment: behaviour fits optimism about future time demands; showing people information on their own past choices made them less inconsistent later. Supports showing personal history; weak until published.
• Emanuel A, Schenk M, Heller AS, Otto AR (2025). Effects of task duration prediction errors on affective state and performance. PsyArXiv preprint, doi:10.31234/osf.io/8usp6_v1 (preprint, not peer-reviewed). Five experiments, n = 628: a task turning out shorter than expected improved mood; longer than expected worsened it; effects on performance inconsistent. If it holds, planning to a conservative time protects mood.
• Reading for Sioul: (a) for aversive tasks the risk is delay and fragmentation (more sessions, restart costs) more than a known estimate bias; (b) for absorbing or wanted tasks, overruns often come from not stopping (felt time shrinks, hyperfocus), which is a choice or a stopping problem, not a misestimate — the app should be able to tell "kept going on purpose" from "took longer"; (c) the person's mood about a task is a plausible kind for separate ratios, but the evidence does not say which way it shifts the ratio, so learn it, do not hard-code it.

3.2 Difficulty and complexity¶

• Thomas KE, Newstead SE, Handley SJ (2003). Exploring the time prediction process: the effects of task experience and complexity on prediction accuracy. Applied Cognitive Psychology 17(6):655–673. doi:10.1002/acp.893 (verified). Simple 3-disk Tower of Hanoi: no underestimation, "participants consistently overestimated"; prior task experience improved accuracy; more complex 4–5-disk versions were less biased; a short, undemanding task was overestimated too. Moderate (four lab experiments).
• Halkjelsvik & Jørgensen 2012 (1.6): "Our summaries challenge earlier results regarding the effects of factors such as complexity/difficulty and experience." No reliable general rule that harder tasks are more underestimated.
• Hinds PJ (1999). The curse of expertise: the effects of expertise and debiasing methods on prediction of novice performance. Journal of Experimental Psychology: Applied 5(2):205–221. doi:10.1037/1076-898X.5.2.205 (verified). Experts underestimated how long novices take and resisted debiasing. Moderate. Relevant if the app ever suggests durations learned from other people or from the person's past self in a better state.
• Kruger & Evans 2004 (1.4): unpacking raised estimates for complex multi-step tasks, not simple ones.

3.3 Familiar versus new¶

• Roy & Christenfeld 2007 (1.3): underestimation disappears when the task is novel (no biased memory to draw on).
• Boltz, Kupperman & Dunne 1998 and the origami study (via Halkjelsvik & Jørgensen 2018, 1.6): with practice, estimates move from too high to too low while accuracy improves; chunking hides the parts of well-known tasks.
• Tobin S, Grondin S (2015). Prior task experience affects temporal prediction and estimation. Frontiers in Psychology 6:916. doi:10.3389/fpsyg.2015.00916 (verified, PMID 26217261). Expert runners predicted and recalled their 5 km time more accurately than intermediates. Weak–moderate.
• Halkjelsvik & Jørgensen 2018: relevance of experience matters, its amount much less.
• Reading for Sioul: new tasks should get a wider uncertainty, not a different multiplier; routine tasks can drift toward underestimation over time as they become automatic, which a moving window will track.

3.4 Short versus long¶

• Roy & Christenfeld 2008 (1.3): within 1–15 min tasks, short (< 2 min) overestimated, longer ones underestimated, for both memory and prediction.
• Burt CDB, Kemp S (1994). Construction of activity duration and time management potential. Applied Cognitive Psychology 8(2):155–168. doi:10.1002/acp.2350080206 (verified). Everyday activities: "subjects generally made overestimations"; expected and remembered durations correlated beyond actual duration (both built from general knowledge); accuracy unrelated to time-management questionnaire scores. Weak–moderate.
• Halkjelsvik & Jørgensen 2018 (4.1): the magnitude effect is real in controlled experiments (small tasks over-, large tasks underestimated relative to context) but exaggerated by statistical artefacts in observational data; small software tasks were unbiased or overestimated, 100 h+ projects about 20 % over.
• Jørgensen M, Halkjelsvik T, Kitchenham B (2012). How does project size affect cost estimation error? Statistical artifacts and methodological challenges. International Journal of Project Management 30(7):839–849. doi:10.1016/j.ijproman.2012.01.007 (verified). Whether larger projects look over- or underestimated depends on whether size is measured by the estimate or the actual (regression to the mean). Strong methodological point.
• Reading for Sioul: one multiplicative ratio for all lengths will over-pad short tasks and under-pad long ones. Model the ratio as a function of the estimate (the only size known at planning time); regression toward the middle given the estimate is a real, usable effect for prediction, not an error to remove.

3.5 Interruptions and fragmentation¶

• Czerwinski M, Horvitz E, Wilhite S (2004). A diary study of task switching and interruptions. CHI 2004:175–182. doi:10.1145/985692.985715 (verified). Week-long diaries of information workers: many interleaved tasks, frequent interruptions, hard returns. Moderate (small diary study).
• Mark G, Gonzalez VM, Harris J (2005). No task left behind? Examining the nature of fragmented work. CHI 2005:321–330. doi:10.1145/1054972.1055017 (verified). 24 observed workers: "57% of their working spheres are interrupted"; most interrupted work resumes the same day but after more than two other activities. Moderate.
• Mark G, Gudith D, Klocke U (2008). The cost of interrupted work: more speed and stress. CHI 2008:107–110. doi:10.1145/1357054.1357072 (verified). Interrupted tasks were finished faster at the price of more stress, frustration and effort. Moderate (lab).
• Reading for Sioul: interruptions are part of a task's real duration (EBS keeps the clock running, 4.5), but a timer that runs through a long interruption measures the interruption, not the task. Record sessions (start/stop), not a single total, so both can be told apart.

3.6 Planning context¶

• Tonietto GN, Malkoc SA, Nowlis SM (2019). When an hour feels shorter: future boundary tasks alter consumption by contracting time. Journal of Consumer Research 45(5):1085–1102 (online 2018). doi:10.1093/jcr/ucy043 (verified). Eight studies: an hour before a scheduled appointment feels shorter than an open hour; people do fewer and shorter things in it; making a long task easy to split into subtasks reduces the effect. Moderate–strong.
• Tonietto GN, Malkoc SA (2016). The calendar mindset: scheduling takes the fun out and puts the work in. Journal of Marketing Research 53(6):922–936. doi:10.1509/jmr.14.0591 (verified). 13 studies: scheduling a leisure activity makes it "less free-flowing and more work-like" and less enjoyed; "roughly scheduling (without prespecified times) eliminates this effect". Strong. A planner should place rest and leisure roughly, not to the minute.
• Zauberman G, Lynch JG (2005). Resource slack and propensity to discount delayed investments of time versus money. Journal of Experimental Psychology: General 134(1):23–37. doi:10.1037/0096-3445.134.1.23 (verified). Seven experiments: "people expect slack for time to be greater in the future than in the present", more so than for money, so they over-commit future time. Moderate–strong. Plans for "next week" will be more over-full than plans for today.
• Peetz J, Buehler R, Wilson A (2010). Planning for the near and distant future: how does temporal distance affect task completion predictions? Journal of Experimental Social Psychology 46(5):709–720. doi:10.1016/j.jesp.2010.03.008 (verified; findings from the authors' thesis summary). Near tasks bring concrete thoughts (more optimism) and thoughts of obstacles (less optimism) at once. Moderate.

4. Correcting estimates: reference classes, personal history, feedback¶

4.1 Halkjelsvik & Jørgensen 2018 — methods chapter (book cited in 1.6)¶

Points from chapters 5–7 of the open-access book, each with the underlying study when the book names it:

• Use the history of your own errors, not min/max guesses. Their method: keep the person's single estimate (whatever it means to them, "as long as you are reasonably consistent"), then read two percentiles of the past error ratio (e.g. "in 8 of 10 similar tasks I spent 150 % or less") and fit a lognormal; any pX follows. Example: estimate 10 h, p20 = 90 %, p80 = 150 % of estimate gives p50 ≈ 11.6 h, p95 ≈ 19 h, and only ~30 % chance of finishing within the 10 h estimate. They suggest p50 as the planned time and p85 as the budget. Opinion/method, built on the lognormal finding and the overconfidence studies below.
• Self-made intervals are too narrow and feedback fixes them only partly. 90 % intervals held 64 % of actuals over the first 10 tasks, 70 % over the next 10, 81 % over the last 10 (Jørgensen & Teigen 2002, conference paper). People doing the work themselves stayed "strongly overconfident"; outsiders reading the same data were accurate (Gruschke & Jørgensen 2008, ACM TOSEM 17(4)). So: learn the spread from data, do not ask the person for a range.
• Anchors are strong and survive warnings. Software professionals told a customer expected 4 h, 40 h or 800 h gave medians of 60 h, 100 h, 300 h; control 160 h (Jørgensen & Sjøberg 2004, IJPM 22(4)). Sequence effect: the same medium task got a median of 195 h after predicting a large task and 95 h after a small one. Anything the app shows before the person types an estimate (a suggested duration, a padded value, the previous task's estimate) will move the estimate.
• Inverted format inflates optimism. "How much can you do in 5 minutes?" gave 0.8 pages/min predicted, "in 30 minutes" gave 0.3 pages/min (Halkjelsvik, Jørgensen & Teigen 2011). 12 errands in 2 h were predicted to fit; they took about 5 h. A planner that asks "what fits in your 2 free hours?" invites this; asking "how long will this take?" is safer.
• Magnitude effect is partly an artefact. If tasks are grouped by actual time, big ones look underestimated; grouped by estimated time, big estimates look overestimated — both from random error alone (13 studies compared: all 7 that sized by actual found more underestimation of big tasks; those sizing by estimate found little or none). Controlled experiments still show a real effect for small tasks: small overestimated, large underestimated, relative to the context (Vierordt, central tendency). The book: adjusting toward the middle of similar tasks "will result in a magnitude bias but may also increase the accuracy".
• Unpacking/decomposition raise estimates; whether that helps depends on the direction of the bias. Checklists: software projects using them were 10 % low on average versus 38 % low without (Furulund & Moløkken-Østvold 2007, IEEE conference paper, as cited by the book; not checked). Office tasks that tended to be overestimated became more overestimated when decomposed (+9–10 % versus +2–5 % whole); tasks that tended to be underestimated became accurate (0–8 % low versus 13–26 % low whole) (Forsyth & Burt 2008). Concrete step-by-step plans can increase optimism for completion dates.
• Analogies, then shrink toward the broader average. Use close past tasks; if they are few or unusual, average them with the broader class; "a simple rule of thumb … is to use a 50–50 weighting of the closest analogy and the average"; more weight to the close analogy when similar tasks vary little. This is informal empirical-Bayes shrinkage.
• Rewarding accuracy backfires: people stretch work to fit a high estimate or cut quality to fit a low one. A planner should not score the person on hitting estimates.
• Time unit effect: coarser units (days) give higher predictions than finer ones (hours) for the same work (74 software professionals, Jørgensen 2016). Keep one unit (minutes) in the app so the ratio is not shifted by unit changes.
• Strength overall: moderate (many findings come from software professionals and students; several rest on one or two studies, as the authors state).
• Underlying references checked on Crossref: Jørgensen & Sjøberg 2004, Int J Project Management 22(4):317–325, doi:10.1016/S0263-7863(03)00085-1; Gruschke & Jørgensen 2008, ACM TOSEM 17(4):1–35, doi:10.1145/13487689.13487693; Halkjelsvik, Jørgensen & Teigen 2011, "To read two pages, I need 5 minutes, but give me 5 minutes and I will read four", Applied Cognitive Psychology 25(2):314–323, doi:10.1002/acp.1693; Jørgensen 2016, "Unit effects in software project effort estimation: work-hours gives lower effort estimates than workdays", J Systems and Software 117:274–281, doi:10.1016/j.jss.2016.03.048; Connolly & Dean 1997, Management Science 43(7):1029–1045, doi:10.1287/mnsc.43.7.1029; Boltz, Kupperman & Dunne 1998, Memory & Cognition 26(5):903–921, doi:10.3758/BF03201172; Jørgensen, Faugli & Gruschke 2007, J Systems and Software 80(9):1472–1482, doi:10.1016/j.jss.2006.09.047; Nan & Harter 2009, IEEE TSE 35(5):624–637, doi:10.1109/TSE.2009.18 (inverted-U between schedule pressure and productivity); Pezzo, Litman & Pezzo 2006, Personality and Individual Differences 41(7):1359–1371, doi:10.1016/j.paid.2006.03.029 (present-focused, carefree people predict early finishes for boring tasks and finish later). Jørgensen & Teigen 2002 (ProMAC conference) not checked — "unverified".

4.2 Feedback and training: the PSP evidence¶

• Prechelt L, Unger B (2001). An experiment measuring the effects of Personal Software Process (PSP) training. IEEE Transactions on Software Engineering 27(5):465–472. doi:10.1109/32.922716 (verified). Controlled comparison of students just out of a PSP course against students from other courses, one programming task. From the technical-report version (Univ. Karlsruhe TR 1/1999, read): PSP estimates were "probably" better "although … the difference is not overly consistent"; misestimates were less variable and the worst underestimate was much smaller (about 3× the median), but "the difference in mean or median is not significant". A few dozen students. Halkjelsvik & Jørgensen count it as "no improvement" in accuracy; fairer: less spread, same average.
• Abrahamsson P, Kautz K (2002). Personal Software Process: classroom experiences from Finland. In Software Quality — ECSQ 2002, LNCS 2349:175–185. doi:10.1007/3-540-47984-8_21 (verified). Classroom PSP data; cited by Halkjelsvik & Jørgensen as no estimation gain. Companion paper: Abrahamsson P, Kautz K, Sieppi H, Lappalainen J (2002). Improving software developer's competence: is the Personal Software Process working? PROFES 2002 workshop, Rovaniemi; arXiv:1311.0228 (read): 58 master's students in Finland and Denmark; "the use of PSP did not improve size and time estimation skills" (SEI's claimed median 1.75× gain in effort estimation "not supported"). Weak–moderate.
• Jørgensen M, Gruschke TM (2009). The impact of lessons-learned sessions on effort estimation and uncertainty assessments. IEEE TSE 35(3):368–383. doi:10.1109/TSE.2009.2 (verified). Reflection sessions on past estimation errors did not improve accuracy beyond what task knowledge alone gave.
• Gruschke & Jørgensen 2008 (above): outcome feedback narrows overconfidence only slowly when the estimator does the work.
• Strength: moderate for "reflection-based training alone does not make people estimate better"; the studies are small (students, tens of developers). Contrast with Roy, Mitten & Christenfeld 2008 (1.3): showing measured durations at the moment of estimating did help. The difference matters for Sioul: an app need not train the person; it can apply the correction itself and show relevant measured history when asked.

4.2b Feedback on durations in the lab: mixed¶

• König CJ, Wirz A, Thomas KE, Weidmann RZ (2015). The effects of previous misestimation of task duration on estimating future task duration. Current Psychology 34(1):1–13. doi:10.1007/s12144-014-9236-3 (verified; abstract via the authors' repository). Two unrelated lab tasks: people who first estimated how long task 1 had taken (self-generated feedback) underestimated task 2 less — "a general kind of learning from previous misestimation". Weak–moderate.
• Thomas KE, König CJ (2018). Knowledge of previous tasks: task similarity influences bias in task duration predictions. Frontiers in Psychology 9:760. doi:10.3389/fpsyg.2018.00760 (verified, PMID 29881362). Three experiments, tasks a week apart: predictions were less biased when the first task was the same task, but "duration feedback did not influence bias". Moderate. Conflicts with Roy, Mitten & Christenfeld 2008 (1.3).
• Thomas KE, Handley SJ, Newstead SE (2007). The role of prior task experience in temporal misestimation. Quarterly Journal of Experimental Psychology 60(2):230–240. doi:10.1080/17470210600785091 (verified, PMID 17455056). Three experiments: predictions anchor on the duration of the preceding task — a longer task before led to overestimating a shorter one, a shorter task before led to underestimating a longer one; "using such information does not necessarily improve accuracy". Moderate.
• Thomas KE, Handley SJ (2008). Anchoring in time estimation. Acta Psychologica 127(1):24–29. doi:10.1016/j.actpsy.2006.12.004 (verified, PMID 17276378). Two experiments: before estimating, people picked a duration figure from earlier participants (same task or a different one); estimates moved toward that figure — underestimation after a shorter anchor, overestimation after a longer one. Moderate. A suggested duration shown in the app is exactly such an anchor.
• Jørgensen M, Teigen KH, Moløkken K (2004). Better sure than safe? Over-confidence in judgement based software development effort prediction intervals. Journal of Systems and Software 70(1–2):79–93. doi:10.1016/S0164-1212(02)00160-7 (verified; not read). Teigen KH, Jørgensen M (2005). When 90% confidence intervals are 50% certain: on the credibility of credible intervals. Applied Cognitive Psychology 19(4):455–475. doi:10.1002/acp.1085 (verified; title states the finding). Both: self-made effort intervals are far too narrow. Moderate.
• Net: what reliably helps is relevant history (same or very similar task) at the moment of predicting; generic feedback ("you underestimate") helps little and inconsistently. For an app this argues for the app doing the correction, and for showing the person the measured times of similar past tasks when they estimate — or not at all, since any number shown is also an anchor (4.1).

4.3 Reference-class forecasting (the "outside view")¶

• Flyvbjerg B (2006). From Nobel Prize to project management: getting risks right. Project Management Journal 37(3):5–15. doi:10.1177/875697280603700302 (verified). Method: forecast from "actual performance in a reference class of comparable projects", bypassing optimism and strategic misrepresentation; first use was cost forecasts for UK transport projects.
• Flyvbjerg B (2008). Curbing optimism bias and strategic misrepresentation in planning: reference class forecasting in practice. European Planning Studies 16(1):3–21. doi:10.1080/09654310701747936 (verified; online 2007). Same method, endorsed by the American Planning Association; applied to the Edinburgh Tram and Crossrail. The uplift is read from the class's overrun distribution at a chosen acceptable risk of overrun.
• HM Treasury (2003; updated 2013). Supplementary Green Book Guidance: Optimism Bias, built on Mott MacDonald (2002) Review of Large Public Procurement in the UK (existence checked via gov.uk and NHS Wales pages; not read in full). Uplifts per project category: upper bounds for works duration of 4 % (standard buildings), 39 % (non-standard buildings), 54 % (equipment/development), to be lowered as the project shows evidence that risks are managed. Practice — a precedent for per-kind uplifts that start high and shrink with evidence.
• Batselier J, Vanhoucke M (2016). Practical application and empirical evaluation of reference class forecasting for project management. Project Management Journal 47(5):36–51. doi:10.1177/875697281604700504 (verified). On real project data, RCF "performs best, for both cost and time forecasting" against Monte Carlo and earned-value methods. Moderate (project-level data, one research group).
• For a single person, the reference class is their own past tasks of the same kind; the evidence that this beats intuition for individuals comes from Roy, Mitten & Christenfeld 2008 (1.3) and the book's analogy chapter (4.1), not from Flyvbjerg's megaprojects.

4.4 Personal Software Process (Humphrey) and PROBE¶

• Humphrey WS (1995). A Discipline for Software Engineering. Addison-Wesley. (Book; cited by the SEI body of knowledge below; not read.) Engineers log time per task, estimate from personal history, and review estimate-versus-actual after each program.
• Pomeroy-Huff M, Cannon R, Chick TA, Mullaney J, Nichols W (2009). The Personal Software Process (PSP) Body of Knowledge, Version 2.0. Software Engineering Institute, CMU/SEI-2009-SR-018 (read). PROBE picks the estimator by data quality:
  • Method A/B (regression of actual on estimate): only with "three or more data points … that correlate", intercept near 0, and slope β1 "between 0.5 and 2" for size, "within 50 % of 1/(historical productivity)" for time; data fit for planning "if the r² value is greater than 0.5 and if the tail area in the significance calculation is ≤ 0.05".
  • Method C (no usable regression, some history): β0 = 0, β1 = actual total to date / planned total to date — a plain ratio of sums.
  • Method D (no history): judgment.
  • Warnings: "Extreme data points can lead to erroneous β0 and β1 values, even with high correlation"; estimates outside the range of past data "are likely to be seriously in error".
  • Established practice; directly analogous to Sioul: a ratio of sums as the default, a fitted slope only when the data justify it, bounds on the slope, no extrapolation.
• Evidence on PSP: Hayes W, Over JW (1997), The Personal Software Process: an empirical study of the impact of PSP on individual engineers, SEI CMU/SEI-97-TR-001, doi:10.21236/ADA335543 (verified): 298 engineers in PSP courses; SEI reports a median 1.75× improvement in effort-estimation accuracy (figure as quoted by Abrahamsson & Kautz) — run by the method's authors, within-course, confounded with growing task familiarity. Independent classroom replications found "no change or improvement" in size or time estimation (Abrahamsson, Kautz, Sieppi & Lappalainen 2002, 58 students; 4.2) and only less spread (Prechelt & Unger 2001; 4.2). Wesslén A (2000), Empirical Software Engineering 5(2):93–123, doi:10.1023/A:1009811222725 (verified; not read) replicated the SEI analysis.
• Data quality: Johnson PM, Disney AM (1998). The personal software process: a cautionary case study. IEEE Software 15(6):85–88. doi:10.1109/52.730851 (verified); and (1999) A critical analysis of PSP data quality: results from a case study. Empirical Software Engineering 4(4):317–349. doi:10.1023/A:1009801218527 (verified). Manually recorded PSP data were error-prone enough to distort the derived measures; they argue for automated collection. Moderate. For Sioul: rely on timer data, treat hand-entered actuals as lower quality, and catch obvious errors at entry.

4.5 Evidence-Based Scheduling (Spolsky, FogBugz)¶

• Spolsky J (2007, 26 Oct). Evidence based scheduling. Joel on Software blog, https://www.joelonsoftware.com/2007/10/26/evidence-based-scheduling/ (read). Practitioner opinion; shipped in FogBugz; no published evaluation found.
  • "When you divide estimate by actual, you get velocity" (so velocity 0.5 means tasks take twice the estimate).
  • Monte Carlo: 100 simulated futures; each future task's estimate is divided by "a randomly-selected velocity from that developer's historical velocities"; the spread of totals gives a probability for each ship date.
  • "Throw away any velocities older than, say, six months."
  • Tasks must be small: "Nothing longer than 16 hours."
  • Interruptions: keep the clock running on the current task, so the velocities absorb typical interruptions.
  • A consistent estimator (velocities all near 0.6) gives a narrow distribution even though biased; an erratic one gives a wide one. Bias is fine; inconsistency is the problem.
  • Limits: samples each task's velocity independently, so it ignores bad days that slow every task at once (Halkjelsvik & Jørgensen 2018 warn that Monte Carlo without dependencies gives falsely narrow totals); six months is far longer than Sioul's 7–15-day window.

4.6 Agile velocity and flow metrics¶

• Beck K, Fowler M (2000). Planning Extreme Programming. Addison-Wesley (book, not read): "yesterday's weather" — plan the next iteration to hold what the last one actually delivered. Cohn M (2005). Agile Estimating and Planning. Prentice Hall (book, not read): velocity as a range from recent iterations. Practice. Halkjelsvik & Jørgensen 2018 note that the "how much fits in this time" format, usually over-optimistic, "seems to lead to fairly accurate predictions" in agile teams because it is anchored on measured past output.
• Kanban practice uses percentiles of past cycle times (e.g. "85 % of items finish within N days") instead of means. Practice, practitioner books; not verified here.

4.7 Parkinson's law, buffers and per-task padding¶

• Parkinson CN (1955 essay; 1957 book Parkinson's Law): "work expands so as to fill the time available for its completion". Satire/opinion, but tested:
• Aronson E, Gerard EW (1966). Beyond Parkinson's law: the effect of excess time on subsequent performance. JPSP 3(3):336–339. doi:10.1037/h0023000 (verified). People given too much time on one task then took longer on a similar self-paced task: work "continues to require more time". Weak (one lab study).
• Bryan JF, Locke EA (1967). Parkinson's law as a goal-setting phenomenon. Organizational Behavior and Human Performance 2(3):258–275. doi:10.1016/0030-5073(67)90021-9 (verified; abstract not available — usually summarised as: more time allowed, easier implicit goals, slower work).
• Brannon LA, Hershberger PJ, Brock TC (1999). Timeless demonstrations of Parkinson's first law. Psychonomic Bulletin & Review 6(1):148–156. doi:10.3758/BF03210823 (verified, PMID 12199310). People told a later task was cancelled "dallied" on the current one; replicated twice exactly and once conceptually. Moderate.
• Trietsch D, Mazmanyan L, Gevorgyan L, Baker KR (2012). Modeling activity times by the Parkinson distribution with a lognormal core: theory and validation. European Journal of Operational Research 216(2):386–396. doi:10.1016/j.ejor.2011.07.054 (verified). Field data from several sources: activity times are lognormal once two distortions are removed — rounding, and the Parkinson effect (early finishes reported as on time, i.e. "random censoring of earliness"); activities show positive stochastic dependence. Moderate–strong. For Sioul: if the padded plan is shown as the target, finished-early tasks will be stretched or logged as on time, biasing the ratio upward.
• Buehler R, Peetz J, Griffin D (2010). Finishing on time: when do predictions influence completion times? OBHDP 111(1):23–32. doi:10.1016/j.obhdp.2009.08.001 (verified; full text not read). Predictions can steer completion; cited by Halkjelsvik & Jørgensen 2018 as evidence that time predictions change the work.
• Goldratt EM (1997). Critical Chain (book): cut each task estimate to about its median and put the safety in one shared buffer at the end, because padding inside tasks is consumed (Parkinson, "student syndrome"). Leach LP (1999), Project Management Journal 30(2):39–51, doi:10.1177/875697289903000207 (verified): practitioner case for it. Raz T, Barnes R, Dvir D (2003), Project Management Journal 34(4):24–32, doi:10.1177/875697280303400404 (verified): critical review — useful ideas, not a complete solution, weak evidence. Opinion/practice; the shared-buffer idea is also what the statistics recommend (5.8: the spread of a sum is smaller than the sum of the spreads).

4.8 Feedback, calibration, and how showing accuracy affects people¶

• Lichtenstein S, Fischhoff B (1980). Training for calibration. Organizational Behavior and Human Performance 26(2):149–171. doi:10.1016/0030-5073(80)90052-5 (verified). Two experiments of intensive training with feedback: "considerable learning almost all of which was accomplished after receipt of the first feedback", with limited generalization (summary via secondary sources). Moderate.
• Moore DA, Healy PJ (2008). The trouble with overconfidence. Psychological Review 115(2):502–517. doi:10.1037/0033-295X.115.2.502 (verified). "Overprecision appears to be more persistent" than over-estimation of oneself. Strong. Asking a person for a range will give a range too narrow; learn spread from data.
• Bolger F, Önkal-Atay D (2004). The effects of feedback on judgmental interval predictions. International Journal of Forecasting 20(1):29–39. doi:10.1016/S0169-2070(03)00009-8 (verified; abstract not available — not summarised here).
• Kluger AN, DeNisi A (1996). The effects of feedback interventions on performance: a historical review, a meta-analysis, and a preliminary feedback intervention theory. Psychological Bulletin 119(2):254–284. doi:10.1037/0033-2909.119.2.254 (verified). 607 effect sizes, 23,663 observations: average d = .41 but "over 1/3 of the FIs decreased performance"; "FI effectiveness decreases as attention moves up the hierarchy closer to the self and away from the task." Strong. A number framed as a trait of the person ("you are bad at estimating") is the kind of feedback that backfires; a number framed as a property of the task and the plan is not.
• Hattie J, Timperley H (2007). The power of feedback. Review of Educational Research 77(1):81–112. doi:10.3102/003465430298487 (verified). Review with a four-level model (task, process, self-regulation, self): feedback about the task and the process helps; feedback about the self as a person (praise or blame) helps least. Strong (education).
• Etkin J (2016). The hidden cost of personal quantification. Journal of Consumer Research 42(6):967–984. doi:10.1093/jcr/ucv095 (verified). Six experiments: measuring an activity raises how much people do but "can simultaneously reduce how much people enjoy those activities … can make enjoyable activities feel more like work", lowering continued engagement and well-being. Moderate–strong. Timing leisure or pleasant tasks has a cost; tracking should be optional there.
• Sirois FM (2014). Procrastination and stress: exploring the role of self-compassion. Self and Identity 13(2):128–145. doi:10.1080/15298868.2013.763404 (verified). Four samples (145, 339, 190, 94): procrastination goes with lower self-compassion and more stress; self-compassion mediates. Moderate (correlational).
• Wohl MJA, Pychyl TA, Bennett SH (2010). I forgive myself, now I can study: how self-forgiveness for procrastinating can reduce future procrastination. Personality and Individual Differences 48(7):803–808. doi:10.1016/j.paid.2010.01.029 (verified; abstract not available here). Students who forgave themselves for procrastinating before one exam procrastinated less before the next (as commonly summarised). Weak–moderate.
• Beaton DM, Sirois F, Milne E (2022). Experiences of criticism in adults with ADHD: a qualitative study. PLOS ONE 17(2):e0263366. doi:10.1371/journal.pone.0263366 (verified, PMID 35180241). 162 adults with ADHD or high traits: inattention-related behaviour is "the most criticised"; criticism harms self-worth; being understood changes whether it is felt as criticism. Moderate (qualitative). A planner that comments on lateness or misestimates lands on a sore spot for this group.
• Li I, Dey A, Forlizzi J (2010). A stage-based model of personal informatics systems. CHI 2010:557–566. doi:10.1145/1753326.1753409 (verified). Barriers cascade from collection to reflection; recommends balancing automation and user control. Epstein DA, Ping A, Fogarty J, Munson SA (2015). A lived informatics model of personal informatics. UbiComp 2015:731–742. doi:10.1145/2750858.2804250 (verified). Surveys (105, 99, 83) and 22 interviews: people lapse and resume tracking; design for it. Moderate (qualitative/survey).
• Kersten-van Dijk ET, Westerink JHDM, Beute F, IJsselsteijn WA (2017). Personal informatics, self-insight, and behavior change: a critical review of current literature. Human–Computer Interaction 32(5–6):268–296. doi:10.1080/07370024.2016.1276456 (verified). Of 6,568 papers only 24 qualified; insight is often reported but evidence that data alone changes behaviour is weak. Strong as a review; weak evidence base. Do not expect the displayed ratio to train the person; let the planner use it.
• Weiser M, Brown JS (1996). Designing calm technology. PowerGrid Journal 1.01 (essay, Xerox PARC; widely reprinted; opinion; not checked against the original venue). Inform at the periphery; move to the centre only when needed. Fits showing the ratio quietly, on request.

4.9 Clinical and tool practice for ADHD and fatigue¶

• Solanto MV, Marks DJ, Wasserstein J, et al. (2010). Efficacy of meta-cognitive therapy for adult ADHD. American Journal of Psychiatry 167(8):958–968. doi:10.1176/appi.ajp.2009.09081123 (verified, PMID 20231319). RCT, 88 adults, 12 weekly group sessions on time management, organisation and planning plus work on discouraging thoughts; better inattention outcomes than supportive therapy (53 % vs 28 % responders, per press summaries). Whether the manual uses estimate-versus-actual exercises was not confirmed here. Strong for the package, not for any single technique.
• Safren SA, Sprich S, Mimiaga MJ, et al. (2010). Cognitive behavioral therapy vs relaxation with educational support for medication-treated adults with ADHD and persistent symptoms. JAMA 304(8):875–880. doi:10.1001/jama.2010.1192 (verified; not read). CBT including organisation and planning modules. Strong for the package.
• Common clinical advice ("multiply your estimate by 1.5 or 2", "externalise time" with visible clocks and timers — Barkley's books) is opinion; no trial of a fixed multiplier was found.
• Tools (product pages, read 2026-10-05; practice, no evaluations):
  • Amazing Marvin: "Compare estimates vs. actual time"; a "Capacity Estimator" that is "Based on your actual pace (learns from your history)" and warns when a day is overfull; formula not published.
  • Sunsama: daily workload threshold with warnings when planned task time exceeds it; optional timer or manual "actual time"; users ask for planned-versus-actual to be reflected on the calendar.
  • goblin.tools "Estimator" (made for neurodivergent users): a language model guesses durations; no personal history.
  • Reclaim/Motion: auto-scheduling with fixed buffers around meetings; no published learning of personal ratios found.
  • FogBugz EBS (4.5) remains the only well-documented tool that learns per-person ratios and plans with percentiles.

5. Statistics for the ratio¶

Notation: for task i, estimate e, tracked time a, log-ratio y = ln(a/e). A ratio of 1.5 is y = 0.41; 0.67 is −0.41: logs make over- and underruns symmetric.

5.1 Shape of the data: skewed, roughly lognormal, heavy right tail¶

• Limpert E, Stahel WA, Abbt M (2001). Log-normal distributions across the sciences: keys and clues. BioScience 51(5):341–352. doi:10.1641/0006-3568(2001)051[0341:LNDATS]2.0.CO;2 (verified). Multiplicative processes give lognormal data; summarise with the geometric mean and a multiplicative spread (×/÷). Strong (methods review).
• Little T (2006). Schedule estimation and uncertainty surrounding the cone of uncertainty. IEEE Software 23(3):48–54. doi:10.1109/MS.2006.82 (verified). Over 100 commercial projects at one company: "estimation accuracy of project duration followed a lognormal distribution, and the uncertainty range was nearly identical throughout the project". Moderate (one company).
• Trietsch et al. 2012 (4.7) and Halkjelsvik & Jørgensen 2018 (1.6): lognormal fits once rounding and the Parkinson effect are handled; tails can be heavier than lognormal (rare huge overruns).
• Consequence: work on y, not on a/e. The arithmetic mean of raw ratios is pulled up by a few big overruns and is not the "typical" ratio.

5.2 Accuracy measures¶

• Kitchenham BA, Pickard LM, MacDonell SG, Shepperd MJ (2001). What accuracy statistics really measure. IEE Proceedings – Software 148(3):81–85. doi:10.1049/ip-sen:20010506 (verified). MMRE and pred(25) only describe the spread and kurtosis of z = estimate/actual; the distribution of z is non-normal; look at its location, skew and spread, e.g. with boxplots. Strong (methods).
• Foss T, Stensrud E, Kitchenham B, Myrtveit I (2003). A simulation study of the model evaluation criterion MMRE. IEEE TSE 29(11):985–995. doi:10.1109/TSE.2003.1245300 (verified). MMRE "does not always select the best model"; do not use it. Strong (simulation).
• Tofallis C (2015). A better measure of relative prediction accuracy for model selection and model estimation. Journal of the Operational Research Society 66(8):1352–1362. doi:10.1057/jors.2014.103 (verified). Mean absolute percentage error systematically favours predictions that are too low; use the log of the accuracy ratio; least squares on it "predicts the geometric mean". Strong (methods).
• Morley SK, Brito TV, Welling DT (2018). Measures of model performance based on the log accuracy ratio. Space Weather 16(1):69–88. doi:10.1002/2017SW001669 (verified). Derives the median symmetric accuracy and a signed bias from the log accuracy ratio, robust and easy to read. Strong (methods).
• Hyndman RJ, Koehler AB (2006). Another look at measures of forecast accuracy. International Journal of Forecasting 22(4):679–688. doi:10.1016/j.ijforecast.2006.03.001 (verified). Percentage errors are asymmetric and unstable near zero. Strong.
• For Sioul: the number to learn is the median (or robust mean) of y; the number to show is exp of it ("about 1.3×"), optionally with the spread as a range ("usually between 0.9× and 2×"). A signed "accuracy %" should not be used: it is asymmetric and reads as a grade.

5.3 Which centre: geometric mean, median, ratio of sums, mean¶

• Median and geometric mean of the ratio coincide under a lognormal and are robust; use them per task ("this task will most likely take 1.3× your guess").
• For sums (a day's load), add means, not medians (Halkjelsvik & Jørgensen 2018, 1.6). Under a lognormal, mean = median × exp(σ²/2): × 1.05 for σ = 0.3, × 1.13 for σ = 0.5, × 1.28 for σ = 0.7.
• The PSP "to-date" ratio (Σ actual / Σ estimate, method C, 4.4) is a time-weighted mean ratio: right for capacity, but one long overrun dominates it. It is a reasonable display figure for "how full did my days really get", less so as the per-task estimator.

5.4 Small samples and outliers¶

• Precision of the learned ratio: the standard error of the mean of y is σ/√n. With σ = 0.5, n = 5 gives a 95 % band of ×/÷ 1.55; n = 10, ×/÷ 1.36; n = 20, ×/÷ 1.24; n = 40, ×/÷ 1.17. With σ = 0.7, n = 10 gives ×/÷ 1.54. (Computed here; σ of 0.3–0.7 is an assumption bracketing the book's example — p20 = 0.9×, p80 = 1.5× gives σ ≈ 0.30 — and the heavier tails of real task logs.) So below about 8–10 tasks a personal ratio is mostly noise and must lean on a prior.
• PROBE's rule of thumb is similar in spirit: no regression with fewer than three correlated points; fall back to the ratio of sums, then to judgment (4.4).
• Robust location: Huber PJ (1964). Robust estimation of a location parameter. Annals of Mathematical Statistics 35(1):73–101. doi:10.1214/aoms/1177703732 (verified) — down-weight, do not drop, far points. Rousseeuw PJ, Croux C (1993). Alternatives to the median absolute deviation. JASA 88(424):1273–1283. doi:10.1080/01621459.1993.10476408 (verified) — robust spread estimators. Leys C, Ley C, Klein O, Bernard P, Licata L (2013). Detecting outliers: do not use standard deviation around the mean, use absolute deviation around the median. Journal of Experimental Social Psychology 49(4):764–766. doi:10.1016/j.jesp.2013.03.013 (verified) — flag points beyond median ± 2.5 × MAD (scaled MAD) as a reasonable default (confirmed via secondary summaries). Strong (methods).
• Outliers in a timer log come in two kinds, and only one is an error: (a) measurement errors — timer left running through lunch or overnight, forgotten start, a task ticked off without tracking; (b) real overruns — the long right tail is the phenomenon itself (5.1). Deleting (b) re-creates the planning fallacy inside the app. So: catch (a) at the source (idle detection, asking at stop time when a session crosses a long gap or is many times the estimate, letting the person trim the session), and treat (b) with robust weights (Huber, or winsorising y at about median ± 3 MAD) rather than deletion.
• Censoring: tasks not finished at the end of the day, or abandoned, have a known only as a lower bound. Dropping them biases the ratio down (only the tasks that finished get counted — the selection effect of 1.6). Kaplan EL, Meier P (1958). Nonparametric estimation from incomplete observations. JASA 53(282):457–481. doi:10.1080/01621459.1958.10501452 (verified) is the standard tool; a simpler rule is to hold a task's ratio until it is completed (summing sessions across days) and to count abandoned tasks separately.
• Hand-entered actuals are retrospective judgments with memory bias (Roy et al., 1.3; Block & Zakay, 2.3) and error-prone (Johnson & Disney, 4.4): weight them less or keep them out of the ratio.

5.5 Per-kind ratios with shrinkage (empirical Bayes / partial pooling)¶

• Efron B, Morris C (1975). Data analysis using Stein's estimator and its generalizations. JASA 70(350):311–319. doi:10.1080/01621459.1975.10479864 (verified); Efron B, Morris C (1977). Stein's paradox in statistics. Scientific American 236(5):119–127. doi:10.1038/scientificamerican0577-119 (verified); Morris CN (1983). Parametric empirical Bayes inference: theory and applications. JASA 78(381):47–55. doi:10.1080/01621459.1983.10477920 (verified); Casella G (1985). An introduction to empirical Bayes data analysis. The American Statistician 39(2):83–87. doi:10.1080/00031305.1985.10479400 (verified). Estimating many group means at once, each shrunk toward the overall mean, beats estimating each alone. Strong (classic statistics).
• Rule: kind k with n_k tasks and mean log-ratio ȳ_k gets μ_k = w·ȳ_k + (1−w)·μ_person, with w = n_k / (n_k + σ²/τ²), where σ² is the within-kind variance of y and τ² the variance of true kind means around the person's mean (both estimated from the person's data, τ² from the spread of kind means minus noise). Examples (computed): σ = 0.5, τ = 0.25 gives w = 0.33 at n = 2, 0.56 at 5, 0.71 at 10, 0.83 at 20; σ = 0.5, τ = 0.15 gives 0.15, 0.31, 0.47, 0.64. A kind needs roughly σ²/τ² tasks (here 4–11) before its own data count as much as the pooled ratio.
• The same structure one level up: the person's ratio shrinks toward a default prior until the person has data. Evidence for the default: across everyday lab tasks predictions were unbiased on average; small tasks overestimated; projects 100 h+ about 20 % over (Halkjelsvik & Jørgensen 2012, 2018). A weak prior centred near 1.0–1.2× with σ ≈ 0.5 is defensible; a fixed "×1.5" or "×2" is not.
• Precedents: Kahneman & Tversky's corrective procedure (move the intuitive estimate toward the class mean in proportion to unpredictability, 1.2); the book's 50–50 rule between closest analogy and broader average (4.1); Green Book uplifts by category, lowered with evidence (4.3).
• Kinds should be few and meaningful (e.g. admin, focused work, chores, errands, social, care); dozens of kinds with two tasks each add noise even with shrinkage.

5.6 Dependence on task length (slope on the estimate)¶

• Model y = α + (β−1)·ln e, i.e. ln a = α + β ln e, rather than a single factor, conditioning on the estimate (the only size known when planning) — Jørgensen, Halkjelsvik & Kitchenham 2012 and the book's magnitude chapter (3.4, 4.1). β < 1 means big estimates run proportionally shorter (noise in the estimates, regression to the mean); β > 1 means big tasks run proportionally longer (central tendency in judgment, Roy & Christenfeld 2008; Jazayeri & Shadlen 2010). Which wins is empirical, per person.
• Guard rails from PROBE (4.4): fit β only with enough spread in estimates and enough points; bound it (PROBE uses 0.5–2 for size; tighter, e.g. 0.7–1.3, is safer for small samples); shrink it toward 1; never extrapolate far outside the range of past estimates.
• β changes slowly (it reflects how the person compresses durations), so it can use a longer history than the level α.

5.7 Window length and weighting¶

• Muth JF (1960). Optimal properties of exponentially weighted forecasts. JASA 55(290):299–306. doi:10.1080/01621459.1960.10482064 (verified). EWMA is the optimal forecast when the true level drifts as a random walk observed with noise; the best smoothing constant depends on the ratio of drift to noise. Strong (classic). Also Roberts SW (1959), Technometrics 1(3):239–250, doi:10.1080/00401706.1959.10489860 (verified); Hunter JS (1986), Journal of Quality Technology 18(4):203–210, doi:10.1080/00224065.1986.11979014 (verified); Gardner ES (2006), Exponential smoothing: the state of the art — Part II, IJF 22(4):637–666, doi:10.1016/j.ijforecast.2006.03.005 (verified).
• What "7–15 days" means in weights (computed here, weights by days elapsed):

So a 7–15-day window corresponds to an EWMA half-life of about 4–5 days or a half-Gaussian σ of about 6–8 days. EWMA reacts more to yesterday and keeps a long thin tail; the half-Gaussian weights the last week more evenly and forgets cleanly after ~3 weeks. For people whose days swing (fatigue, ADHD), the half-Gaussian is less jumpy after one bad day; EWMA catches a real change (new medication, relapse) faster. No study compares them for personal task logs — this is a judgment call.

• Weight by days elapsed, not by task count, so a busy day does not crowd out the week; but also require a minimum effective number of tasks (e.g. 8–10). If the window holds fewer, widen it (adaptive bandwidth) or fall back up the shrinkage hierarchy, rather than trust three points.
• Estimate the spread σ and the slope β over a longer history (e.g. 60–90 days) than the level: spread needs more data and moves less.

5.8 Planning with uncertainty: percentiles, sums, days¶

• Pick a percentile, not the mean, for the plan: Halkjelsvik & Jørgensen 2018 suggest p50 as the planned time and p85 as the budget (4.1); Flyvbjerg reads the uplift at an accepted risk of overrun (4.3); EBS shows the whole distribution (4.5).
• Which percentile is a value choice, and the newsvendor model makes it explicit: Qin Y, Wang R, Vakharia AJ, Chen Y, Seref MMH (2011). The newsvendor problem: review and directions for future research. European Journal of Operational Research 213(2):361–374. doi:10.1016/j.ejor.2010.11.024 (verified). The optimal quantity is the quantile c_u/(c_u + c_o), with c_u the cost of reserving too little (overrun: stress, spill-over, shame) and c_o the cost of reserving too much (idle slack). If an overrun costs 3× as much as slack, plan at P75; 4× gives P80; 9× gives P90. For a calm planner aimed at exhausted people, overruns are costly and slack is cheap, which argues for P80–P90 at the day level.
• Percentile multipliers over the median (lognormal; computed): σ = 0.3: P70 1.17, P80 1.29, P90 1.47. σ = 0.5: P70 1.30, P80 1.52, P90 1.90. σ = 0.7: P70 1.44, P80 1.80, P90 2.45.
• With few data, widen: use the predictive distribution (t-quantile and √(1 + 1/n) on the log scale) rather than plugging in σ̂ — Meeker WQ, Hahn GJ, Escobar LA (2017). Statistical Intervals: A Guide for Practitioners and Researchers, 2nd ed. Wiley. doi:10.1002/9781118594841 (verified).
• Sums: padding each task to its own P80 and adding over-reserves when tasks are independent, and the sum of medians under-reserves. Simulation (computed here): five 30-min tasks, median ratio 1.2, σ = 0.5. Sum of per-task medians = 180 min; day median ≈ 190–198; day mean = 204; day P80 = 242 (independent), 255 (within-day correlation 0.3), 262 (0.5); sum of per-task P80s = 274. So compute the day's percentile from the distribution of the sum (Monte Carlo, or the Fenton–Wilkinson lognormal approximation: Fenton LF (1960), IRE Transactions on Communications Systems 8(1):57–67, doi:10.1109/TCOM.1960.1097606, verified), including a shared day effect.
• Dependence matters: bad days slow every task (Trietsch et al. 2012 found positive dependence between activities; Halkjelsvik & Jørgensen warn that Monte Carlo without dependencies is falsely narrow). Model y = μ_kind + d_day + ε, estimate the day-effect variance from history, and resample whole days (block bootstrap; Efron B (1979), Bootstrap methods: another look at the jackknife, Annals of Statistics 7(1):1–26, doi:10.1214/aos/1176344552, verified) if using EBS-style resampling.
• Within-day update: once the first tasks of the day are done, their y values estimate today's d_day (shrunk toward 0); the rest of the plan can be re-weighed by noon. This is the practical payoff of modelling the day effect. (Design proposal; no study of it in personal planners.)
• Where to put the safety margin: per-task display at the median (or the person's own estimate), with the extra reserve held as unallocated slack in the day, follows both the statistics (risk pooling) and the Critical Chain argument against padding that gets consumed (4.7).

Design criteria¶

Numbers in brackets point to the sections above. "Computed" means worked out here from the stated assumptions, not taken from a study.

What to record

1. Store the person's raw estimate in minutes, when it was typed, and whether any suggested or corrected duration was visible at that moment, because anything shown before the estimate anchors it — Jørgensen & Sjøberg 2004; Thomas & Handley 2008; Thomas, Handley & Newstead 2007 [4.1, 4.2b].
2. Store tracked time as timer sessions (start, stop) rather than one total, so idle gaps, interruptions and work spread over several days can be told apart — Mark et al. 2005; Czerwinski et al. 2004; Spolsky 2007 [3.5, 4.5].
3. Store the outcome (done, still open, abandoned) so unfinished work is not silently dropped — Kaplan & Meier 1958; Halkjelsvik & Jørgensen 2018 on selection [5.4, 1.6].
4. Mark whether the actual time came from the timer or was typed afterwards, since typed-in durations are memories with known bias and frequent errors — Roy et al. 2005; Block & Zakay 1997; Johnson & Disney 1999 [1.3, 2.3, 4.4].
5. Store one of a few kinds per task, and optionally two cheap flags — "new to me" and "I chose to keep going" — because novelty widens uncertainty and absorbing tasks overrun by choice rather than by misestimate — Roy & Christenfeld 2007; Thomas et al. 2003; Gable & Poole 2012; Hupfeld et al. 2019 [3.1, 3.3].
6. Keep the estimate's granularity stable (one unit, whole tasks or subtasks, not both mixed), since unpacking, segmentation and time units all shift estimates — Kruger & Evans 2004; Forsyth & Burt 2008; Jørgensen 2016 [1.4, 4.1].

The estimator

1. Learn on the log-ratio ln(actual / estimate) and report exp of its robust centre; never average raw ratios or percentage errors — Limpert et al. 2001; Kitchenham et al. 2001; Foss et al. 2003; Tofallis 2015; Morley et al. 2018 [5.1, 5.2].
2. Weight past tasks by days elapsed with an EWMA half-life of about 4–5 days or a half-Gaussian of σ ≈ 6–8 days, which is what a 7–15-day effective window means — Muth 1960; computed table [5.7].
3. Require about 8–10 effective tasks before the personal ratio stands on its own, and below that shrink it toward a weak default (median about 1.0–1.2×, σ about 0.5) or widen the window — computed σ/√n; Halkjelsvik & Jørgensen 2012, 2018; PSP PROBE minimums [5.4, 5.5, 4.4].
4. Estimate per-kind ratios by partial pooling toward the person's overall ratio with weight n / (n + σ²/τ²) fitted from the person's own data, and keep kinds few — Efron & Morris 1975; Morris 1983; Kahneman & Tversky 1979; Halkjelsvik & Jørgensen 2018 [5.5].
5. Let the ratio depend on the size of the estimate through a log–log slope fitted on a longer history (60–90 days), bounded (about 0.7–1.3), shrunk toward 1 and never extrapolated beyond past estimates — Roy & Christenfeld 2008; Jørgensen, Halkjelsvik & Kitchenham 2012; Pomeroy-Huff et al. 2009 [3.4, 5.6].
6. Estimate the spread and the day-to-day variance over the longer history with robust statistics (MAD) — Rousseeuw & Croux 1993; Trietsch et al. 2012 [5.4, 5.8].
7. Catch timer mistakes at the source (idle detection; a one-tap question at stop time when a session spans a long gap or runs several times the estimate) and down-weight remaining extremes with Huber weights or winsorising near ±3 MAD, never deleting real overruns — Leys et al. 2013; Huber 1964; Halkjelsvik & Jørgensen 2018 [5.4].
8. Count a task's ratio only when it is finished, summing all its sessions, and track abandoned tasks separately rather than ignoring them — Kaplan & Meier 1958; Halkjelsvik & Jørgensen 2018 [5.4].
9. Leave typed-in actual times out of the ratio, or give them low weight — Roy et al. 2005; Johnson & Disney 1999 [5.4].

How the planner uses it

1. Correct the person's own estimate rather than replace it (estimates are biased but informative, r = .77 in Buehler et al. 1994), show each task at its median corrected length, and keep the safety margin as unassigned slack for the whole day, sized from the distribution of the day's total — Buehler et al. 1994; Halkjelsvik & Jørgensen 2018; Raz et al. 2003; computed simulation [1.1, 4.7, 5.8].
2. Size the day's slack at a chosen percentile of the day total, by default P80–P90 for people who pay heavily for overruns, and let the person move it — Qin et al. 2011 (newsvendor); Halkjelsvik & Jørgensen 2018 (p85 budget); Flyvbjerg 2006; Goudsmit et al. 2012 [5.8, 2.2].
3. Compute the day total by simulation (or Fenton–Wilkinson) including a shared "how today is going" effect, never by adding medians or per-task percentiles — Fenton 1960; Trietsch et al. 2012; computed simulation [5.8].
4. Update the rest of today's plan after the first tasks finish, using their ratios as evidence about today — Trietsch et al. 2012; design proposal [5.8].
5. Give new tasks and thin data a higher percentile (wider predictive interval), not a bigger multiplier — Roy & Christenfeld 2007; Thomas et al. 2003; Meeker et al. 2017 [3.3, 5.8].
6. Ask "how long will this take?" and avoid "what can you fit in this gap?", which inflates optimism — Halkjelsvik, Jørgensen & Teigen 2011 [4.1].
7. Never ask for best/worst-case ranges; learn the spread from the record — Connolly & Dean 1997; Jørgensen, Teigen & Moløkken 2004; Moore & Healy 2008; Gruschke & Jørgensen 2008 [4.1, 4.8].
8. Leave unscheduled time unscheduled and place rest and leisure roughly, not to the minute — Tonietto & Malkoc 2016; Tonietto et al. 2019; Zauberman & Lynch 2005 [3.6].
9. Use the corrected durations for capacity, but do not present the padded figure as a countdown or target, since spare time gets absorbed — Brannon et al. 1999; Aronson & Gerard 1966; Bryan & Locke 1967; Trietsch et al. 2012 [4.7].

How to show it

1. Present the ratio as a property of the plan and of task kinds ("tasks like this usually take about 1.3× the first guess; the plan already allows for it"), not as a score of the person — Kluger & DeNisi 1996; Hattie & Timperley 2007 [4.8].
2. Show it on request, quietly, as a typical value with a plain range ("usually between 0.9× and 2×"), with a line that such gaps are ordinary for everyone (long tasks tend to run over, very short ones under) — Buehler et al. 1994; Halkjelsvik & Jørgensen 2012; Roy & Christenfeld 2008; Weiser & Brown 1996 [1.1, 1.3, 1.6, 4.8].
3. When the person is estimating and asks for help, show measured times of the same or very similar past tasks rather than a generic multiplier — Roy, Mitten & Christenfeld 2008; Thomas & König 2018 [1.3, 4.2b].
4. Make timing optional for leisure and pleasant activities — Etkin 2016; Tonietto & Malkoc 2016 [4.8, 3.6].
5. Tolerate gaps in tracking: a week with no data should leave the last estimate in place (with wider uncertainty), not reset it or nag — Epstein et al. 2015; Li et al. 2010 [4.8].

What never to do

1. Never apply a diagnosis-based multiplier (for ADHD, depression, burnout) or infer a condition from the ratio — Ptáček et al. 2022; Mette 2023; Marx et al. 2022; Thönes & Oberfeld 2015 [2.1, 2.2].
2. Never score, colour-code, reward or streak estimation accuracy, and never flag individual tasks as "late" — Halkjelsvik & Jørgensen 2018; Kluger & DeNisi 1996; Beaton et al. 2022; Sirois 2014 [4.1, 4.8].
3. Never write the corrected value back as the person's estimate or pre-fill the estimate field with it, which would make the ratio measure the app instead of the person — Thomas & Handley 2008; Jørgensen & Sjøberg 2004 [4.1].
4. Never drop overruns as "outliers", never learn only from tasks that finished, and never add medians to plan a day — Halkjelsvik & Jørgensen 2018; Trietsch et al. 2012 [5.3, 5.4, 5.8].
5. Never compare the person with other people or with a norm — Kluger & DeNisi 1996; Hattie & Timperley 2007 [4.8].

Open questions / conflicts¶

• Does affect change the direction of the error? Theory says people are most optimistic about unpleasant tasks (Brunnermeier et al. 2017) and that wanting a task done makes predictions optimistic (Buehler et al. 1997); lab work says fun makes time feel shorter (Gable & Poole 2012; Tonietto et al. 2022); a small real-world study found no overall optimistic bias in study plans (Pychyl et al. 2000). No study measured actual/estimate ratios for real tasks split by enjoyment. Sioul's own data, with an optional enjoyment flag, would be new evidence.
• Does feedback improve estimates? Measured past durations shown at prediction time helped (Roy, Mitten & Christenfeld 2008; König et al. 2015); generic duration feedback did not (Thomas & König 2018); reflection-based training did not (Jørgensen & Gruschke 2009; PSP replications). Unclear which applies to an app view people open once a week.
• Parkinson versus externalised time. Showing padded durations may slow people (Brannon et al. 1999; Trietsch et al. 2012), yet ADHD advice is to make time visible (Barkley, opinion), and finishing earlier than expected may lift mood (Emanuel et al. 2025, preprint). Showing elapsed time and the person's own estimate, with padding held as day slack, is a compromise, not a tested solution.
• Window length. 7–15 days assumes the ratio drifts within weeks (fatigue, treatment, season). No study measures how fast a person's estimation ratio drifts. Jørgensen, Faugli & Gruschke 2007 found past optimism the best single predictor of future optimism, but with much noise. EWMA versus half-Gaussian is untested for this use; logging both on real data and scoring next-day predictions (log-score or pinball loss) would decide.
• Magnitude effect: real or artefact? Controlled experiments show short tasks over- and long tasks underestimated (Roy & Christenfeld 2008; Thomas et al. 2003), but observational slopes are confounded by regression to the mean (Jørgensen, Halkjelsvik & Kitchenham 2012). With a few dozen tasks, a fitted slope may add more noise than it removes; start with β = 1 and let it move only with clear evidence.
• Should interruptions count? EBS keeps the clock running so the ratio forecasts elapsed time (Spolsky 2007); for a burnout planner, the person may want "work time" and "elapsed time" apart. Sessions allow both; which one drives the plan is a product decision.
• Double correction. Once people learn their factor, some will pad their own estimates; the ratio then falls toward or below 1 and the window catches up, but for a week the plan double-pads. Detectable (a step change in the ratio) but not studied.
• Which percentile. Overrun and slack costs differ by person and by day; the newsvendor framing is clean, the numbers are not. Default high, let the person adjust.
• ADHD and fatigue evidence is thin. Timing deficits in ADHD are measured over seconds; adult findings are mixed; no planning-fallacy study in ADHD, burnout or ME/CFS was found. Everything specific to these groups here is extrapolation.
• Tracking burden. Measuring costs enjoyment (Etkin 2016) and attention; people lapse (Epstein et al. 2015). How much tracking an exhausted person will sustain, and how the estimator behaves on sparse, biased samples (people track good days more?), is unknown.
• Preprints and working papers. Emanuel et al. 2025 and Breig et al. 2020 are not peer-reviewed; treat as hints.

Status: complete (2026-10-05). All references above were checked on Crossref, PubMed, OpenAlex or the publisher unless marked "unverified" or "not checked".