Executive Function
Probability Matching Task
Also called: binary prediction task, probability learning task, two-choice prediction task, light-guessing task, maximizing versus matching task
Measure whether people always predict the likelier of two lights or spread their predictions to match how often each comes on, scored as the share of likelier-light predictions.
6-9 minUntimed
Free to start: 10 free participants per account, then $0.99 per participant for academic research, or any yearly plan. Try it here first, with no account.
- Papers, 2021 to 2025
- 64
- Evidence grade
- D
- Registered clinical studies
- 0
What it measures
Measures whether people predict the likelier of two outcomes every time or spread their predictions to match how often each occurs. Two lights sit side by side, one coming on more often, and participants predict which will come on, seeing the answer each time. The main score is the share of predictions for the likelier light over the second half of 200 rounds.
Constructs
- Probability learning
- Choice under uncertainty
- Choice strategy
- Learning from feedback
Research uses
- Research on probability learning and choice strategies
- Decision-making and judgment research
- Within-study comparison of learning from outcome feedback
How the paradigm works
The probability matching task measures how people predict a repeated event when one of two outcomes is more likely. Always predicting the likelier outcome, maximizing, is correct most often, yet in hundreds of experiments people have spread their predictions so that each outcome is predicted about as often as it occurs, probability matching (Newell and colleagues, 2013). Large payments, meaningful feedback and long practice moved many participants toward maximizing (Shanks, Tunney and McCarthy, 2002), and in a trial-by-trial task maximizing was chosen half as often as matching and went with higher cognitive ability (West and Stanovich, 2003). Accounts include a search for patterns in a random sequence (Gaissmaier and Schooler, 2008), an intuition that deliberation can override (Koehler and James, 2009), and exploration and recency during learning (Feher da Silva and colleagues, 2017).
More about the paradigm
Two identical lights drawn by code sit side by side. Before each round the participant predicts which light will come on, by click, tap, or the F and J keys, with no time limit. One light, on a side drawn for each session, comes on in 70% of rounds (60% to 90% in the settings), each round drawn independently from the session's seed, as in Feher da Silva and colleagues' task; nothing on screen says how often. The light that came on is lit for 800 ms with a line saying whether the prediction was correct, then a 300 ms blank. Correct predictions earn 1 point each (0 to 10 in the settings; 0 runs without points). There are 3 practice predictions and 200 rounds (100 to 400 in the settings). About 6 to 9 minutes.
The main score is the share of predictions of the likelier light over the second half of the rounds: 1 means always predicting the likelier light, and 0.7 at the standard setting matches the light's chance. The maximizing index rescales it so that matching is 0 and maximizing is 1. The share over the first half shows learning, and the share of predictions that repeat the last light shows how much the latest outcome guides the next prediction. Predicting the less likely light more often than chance over the run invalidates the session. Scores support research and do not diagnose.
Scores
- Likelier Light, Second Half%
- Share of predictions of the likelier light over the second half of the planned rounds.
- 1 is always the likelier light; the light's set chance (0.7 standard) is matching
- Maximizing Indexscore
- The second-half share rescaled by the likelier light's set chance p: (share - p) / (1 - p).
- 0 is matching, 1 is maximizing; below 0 predicts the likelier light less often than it comes on
- Likelier Light, First Half%
- Share of predictions of the likelier light over the first half of the planned rounds.
- Compare with the second half for learning
- Correct Predictions%
- Share of predictions that named the light that came on, over every round.
- Informational
- Following the Last Light%
- Share of predictions that named the light that came on in the round before, practice included.
- Higher means the latest light guided the next prediction more
- Likelier Light Shown%
- Share of rounds in which the likelier light came on, near its set chance.
- Informational
- Points Wonpoints
- Points from correct predictions in the rounds; 0 when the task runs without points.
- Informational
- Median Decision Timems
- Median time to a prediction, at or after 100 ms.
- Informational
- Anticipationscount
- Presses within 100 ms of a round appearing. They are not taken as predictions.
- Quality check
- Rounds Predictedcount
- Predictions in the scored run.
- Fewer than planned is flagged for review
- Planned Roundscount
- Rounds set for the session.
- Informational
Use and evidence
64 papers from 2021 to 2025 name this task in their title or abstract, down 12% compared with 2016 to 2020. That ranks it 277 of 495 test families in our census. See the counts for every test.
Show the numbers
| Year | Papers |
|---|---|
| 2025 | 12 |
| 2024 | 13 |
| 2023 | 9 |
| 2022 | 10 |
| 2021 | 20 |
| 2020 | 14 |
| 2019 | 17 |
| 2018 | 12 |
| 2017 | 15 |
| 2016 | 15 |
| 2015 | 18 |
| 2014 | 27 |
| 2013 | 10 |
| 2012 | 17 |
| 2011 | 16 |
| 2010 | 18 |
| 2009 | 14 |
| 2008 | 10 |
| 2007 | 7 |
| 2006 | 11 |
| 2005 | 8 |
| 2004 | 8 |
| 2003 | 3 |
| 2002 | 7 |
| 2001 | 1 |
| 2000 | 7 |
Use grade D
30 to 99 papers from 2021 to 2025.
Evidence grade D
Little published measurement evidence found.
- Papers on test-retest reliability
- 0
- Papers on reliability or validity
- 27
- Registered clinical studies using it as an outcome
- 0
Note on the counts: includes animal studies.
How the grades work
Use: A is 1,000 or more papers from 2021 to 2025, B 300 to 999, C 100 to 299, D 30 to 99 and E fewer than 30. Evidence: A means at least 25 papers report its test-retest reliability and at least 25 registered clinical studies use it as an outcome; B, at least 8 test-retest papers, and either 5 registered clinical studies or 150 reliability or validity papers; C, at least 2 test-retest papers, 30 reliability or validity papers or 3 registered clinical studies; D, little published measurement evidence found. The counts measure use. They include every paper that names the task in its title or abstract. Grades describe the published research on this task, not ConductCognition's version of it.
Source: ConductCognition test census 0.22.0, counted 3 November 2026 from OpenAlex and ClinicalTrials.gov.
Task design
The trial sequence as the test runs in ConductCognition.
- Lights
- Two identical round lights drawn by code sit side by side. One of them, on a side drawn from the session's seed, comes on in 70% of rounds (standard; 60% to 90% as a setting), each round an independent draw from the seed. Nothing on screen says how often. The whole sequence is saved on every row.
- Each round
- The participant predicts which light will come on by click, tap, or the F and J keys, with no time limit. The light that came on is lit, with a line saying whether the prediction was correct, for 800 ms, then a 300 ms blank. Presses under 100 ms are not taken.
- Points
- 1 point for each correct prediction (standard; 0 to 10, where 0 runs without points), with a line saying that the points are not exchanged for money, or the paid line when the study pays for points.
- Rounds
- 200 (standard; 100 to 400), after 3 practice predictions of the same lights.
- Scores
- The share of predictions of the likelier light over the second half of the rounds, the maximizing index (0 at matching, 1 at maximizing), the same share over the first half, the share of correct predictions and the share that repeat the last light.
Feher da Silva and colleagues (2017) used two squares and a ball over 300 trials, with course credit or money for correct predictions. This test uses lights of its own, points that are not money unless the study pays for them, and a new sequence for each session, so its rates are not directly comparable with published ones. One person's share varies by chance from round to round, so compare groups or conditions. The scores support research and do not diagnose.
Quality checks
Each session is marked valid, needs review or invalid, with the reason, before you see its scores.
Scored rows: probability_matching_round from the scored run, one prediction of the left or right light per round with no time limit. A prediction counts as right when it names the likelier light (chance 50%), whichever light then comes on. Accuracy below chance invalidates and accuracy not above chance is flagged. Predicting one side on nearly every round is not flagged, because the likelier light stays on one side and always predicting it is the best strategy. Answers under 200 ms on at least half the rounds invalidate. A run short of its planned rounds is flagged for review and withholds the percentile.
Why these rules
Accuracy significantly below chance (one-sided exact test, p < .05) cannot come from guessing; it means the response mapping was reversed or the task was not followed (Slick, Sherman and Iverson 1999 treat below-chance forced-choice performance as definite response bias). Chance-level accuracy can be a real result in impaired participants, so it is only flagged. Predicting the less likely light on most rounds earns fewer points under any strategy, so it means the task was not followed. An answer under 200 ms is faster than a simple visual reaction, so it cannot reflect the item. Following response-time effort screening for untimed tests (Wise and Kong 2005), the session is invalid only when at least half the answers are that fast. Early predictions are guesses by design, so accuracy is judged over the whole run.
Running this test
Settings you can change
| Setting | Standard | Allowed |
|---|---|---|
| Rounds | 200 | 100 to 400 |
| Chance of the likelier light (%) | 70 | 60 to 90 |
| Points per correct prediction | 1 | 0 to 10 |
| Payment line | Hypothetical points (standard) | Hypothetical points (standard), The study pays for points |
Yearly plans can change settings within these limits. Free and pay-as-you-go studies use the standard settings.
Need a setting this test does not have?
Devices and languages
- Untimed test: runs on desktops, laptops, tablets and phones. Participant devices
- Languages: English, Spanish, French, German, Italian, Portuguese (Brazil), Chinese (Simplified) and Japanese.
- Available on: Every plan, including the free plan.
- Recruiting online? See the setup guides for Prolific, Sona Systems, CloudResearch Connect and Qualtrics.
Cite and describe
Participants completed the Probability Matching Task in a web browser (ConductCognition task version 1.0.0, scoring version 1.0.0; Shanks, Tunney & McCarthy, 2002; Koehler & James, 2009; Feher da Silva et al., 2017), with the standard settings (rounds: 200; chance of the likelier light (%): 70; points per correct prediction: 1; payment line: Hypothetical points). The main score was likelier Light, second half (share of predictions of the likelier light over the second half of the planned rounds).
ConductCognition versions: task version 1.0.0, scoring version 1.0.0. Every export records the versions each session used, for a methods section.
- Shanks, D. R., Tunney, R. J., & McCarthy, J. D. (2002). A re-examination of probability matching and rational choice. Journal of Behavioral Decision Making, 15(3), 233-250. https://doi.org/10.1002/bdm.413Large payments, meaningful feedback and long practice moved many participants from matching toward always choosing the likelier outcome.
- West, R. F., & Stanovich, K. E. (2003). Is probability matching smart? Associations between probabilistic choices and cognitive ability. Memory & Cognition, 31(2), 243-251. https://doi.org/10.3758/BF03194383In a trial-by-trial task, matching was chosen twice as often as maximizing, and maximizers scored higher in cognitive ability.
- Gaissmaier, W., & Schooler, L. J. (2008). The smart potential behind probability matching. Cognition, 109(3), 416-422. https://doi.org/10.1016/j.cognition.2008.09.007Matching as the outcome of a search for patterns, or of a win-stay, lose-shift shortcut.
- Koehler, D. J., & James, G. (2009). Probability matching in choice under uncertainty: Intuition versus deliberation. Cognition, 113(1), 123-127. https://doi.org/10.1016/j.cognition.2009.07.003Matching that persisted when no pattern could be exploited, while many matchers rated maximizing as better once it was described.
- Feher da Silva, C., Victorino, C. G., Caticha, N., & Baldo, M. V. C. (2017). Exploration and recency as the main proximate causes of probability matching: a reinforcement learning analysis. Scientific Reports, 7, 15326. https://doi.org/10.1038/s41598-017-15587-zThe design: two places on screen, a likelier side at 0.7 drawn independently on every trial, three practice trials and no word about how outcomes were generated.
Source papers
- Feher da Silva C, Victorino CG, Caticha N, Baldo MVC (2017). Exploration and recency as the main proximate causes of probability matching: a reinforcement learning analysis. Scientific Reports, 7:15326.N 84, Young adults
Questions
How long does the Probability Matching Task take?
About 6 to 9 minutes, as listed in the test library.
What device does a participant need?
Untimed test: runs on desktops, laptops, tablets and phones. The browser must be Chrome, Edge, Firefox or Safari.
What scores does it produce?
Likelier Light, Second Half, Maximizing Index, Likelier Light, First Half, Correct Predictions, Following the Last Light, Likelier Light Shown, Points Won, Median Decision Time, Anticipations, Rounds Predicted and Planned Rounds. Each session is also marked valid, needs review or invalid, with the reason.
How widely is the Probability Matching Task used in research?
64 papers from 2021 to 2025 name it in their title or abstract, down 12% compared with 2016 to 2020. That ranks it 277 of 495 test families in the ConductCognition test census (use grade D). Source: ConductCognition test census 0.22.0, counted 3 November 2026 from OpenAlex and ClinicalTrials.gov.
How much reliability evidence does the Probability Matching Task have?
In the published research, 0 papers report its test-retest reliability and 0 registered clinical studies use it as an outcome (evidence grade D in the ConductCognition test census).
Which settings can a researcher change?
Rounds (100 to 400), chance of the likelier light (%) (60 to 90), points per correct prediction (0 to 10) and payment line (Hypothetical points (standard), The study pays for points). Settings can be changed on yearly plans; free and pay-as-you-go studies use the standard settings.
Which languages can participants use?
Participants can take this test in English, Spanish, French, German, Italian, Portuguese (Brazil), Chinese (Simplified) and Japanese. English is on every plan, and the other languages come with yearly plans. Have a native speaker read the machine-translated text before launch.
Is this a clinical test?
No. ConductCognition is for research use only. It runs, scores and exports the test; clinical interpretation is outside the platform.
Every test: the full test list.
Run this test in your next study
Start free with 10 free participants on any test, then use participant credits or a yearly plan.