↓Skip to main content

Tilt Clock

When should you stop playing?

Platform

Enter a Chess.com username to analyse that player's games.

How this works #

Tilt Clock reads your public Chess.com game history and tests the patterns people assume are there: that you play worse the longer a session runs, worse after losing, worse late at night, worse on certain days. For each one it asks whether your own results actually support it.

Often they do not, and it says so.

What the numbers mean #

Score. A win is 1, a draw is ½, a loss is 0. A score rate is the mean across a set of games, so 50% is an even record. It is not the same as a win rate, which ignores draws.

Your baseline. Your own score rate across exactly the games shown, under exactly the filters you have set. Never a population average and never another player’s. When the filters change, the baseline changes with them, so every comparison is against yourself under the same conditions.

A session. A run of games where the next one began less than 60 minutes after the last one ended. An hour is long enough that a break for tea stays inside one sitting, and short enough that tonight and tomorrow night never merge into one. Start times come from each game’s PGN.

Consecutive losses. Counted within a session only. Losing three last night says nothing about how you start today, and carrying the count across the gap would manufacture tilt nobody experienced. A draw breaks a streak.

The bar a pattern has to clear #

  • A group of games is only drawn at 20 games or more. Below that it appears as a hollow marker with no interval, because a bar drawn on 3 games invites a conclusion the data cannot support.
  • A group can only carry the verdict at 30 games or more.
  • Every group gets a 95% Wilson interval, shown as an error bar. Wilson rather than the usual normal approximation, because these are proportions on modest samples near the extremes, where the normal interval will happily claim a range below 0% or above 100%.
  • A difference is called real only when the interval excludes your baseline and the gap is at least 5 percentage points.

The confidence level is also widened for the number of groups tested. Around twenty-three rules and groups are checked. At a plain 95% level that produces roughly one false alarm every time, simply by construction, because testing twenty-three things and reporting the most extreme is not the same as testing one. On synthetic data with no pattern in it at all, the uncorrected version produced a confident verdict on five runs out of twelve. Correcting for the number of comparisons took that to zero. Each view’s own note gives the exact threshold used on your data.

Why “stop after N” rather than “is game 5 unusual” #

The session and losses views are judged by stopping rules rather than bucket by bucket, and that choice does most of the work.

Asking whether your fifth game differs from your baseline splits a history into eight thin slices and tests each one separately. Every slice is small, and every test pays the multiple-comparison cost. Measured against a known 8 percentage point late-session decline, that approach found it once in fifteen tries. It was close to blind to the thing it was built to detect.

Asking instead whether everything from game five onwards is worse than everything before it compares two large groups in a single test. Same data, same threshold, and it finds the same effect sixteen times in twenty, while still reporting nothing at all on data with no pattern in it. It is also closer to the question you have: whether to stop after four games, rather than whether your fifth game is unusual.

The rule chosen is the one where results most change, not the earliest one that technically separates. On a player who wins their first three games and loses after that, “stop after one game” does separate, because everything from game two onwards is worse than game one alone. It is still the wrong advice.

The per-game charts are shown underneath, because they show the shape the rule came from.

What could fool this #

Named here rather than buried, because they are real and this tool cannot rule them out for you.

Rating drift. If you improved over the period and your playing hours shifted over the same period, that appears as a time-of-day effect when it is really just you getting better. Narrow the date range and see whether the effect survives.

Opponent strength. A score rate ignores who you played. If your late-session games happen to be against stronger opponents, that looks exactly like tilt. Every table shows the mean opponent rating per group, so check it is flat before believing anything else in the row.

How rating change is worked out. Chess.com does not publish a rating change per game, only your rating at each game. That figure is the rating after the game, so a game’s change is derived as its rating minus the rating of your previous game in the same time class. Any movement from another cause between two games is attributed to the later one.

Clock use is not examined. The tool reads how each game finished, not how you spent your time inside it. A rising share of losses on time late in a session is visible here. Whether you were moving too fast is not.

Privacy #

Games come from the public Chess.com API through a small caching service. All of the analysis happens in your browser. Nothing about you is stored, there is no account and nothing to sign in to, and no usernames are written to the service logs. Request paths are recorded with the username replaced, so the logs show what kind of request was made and nothing about who made it.