You've lost three in a row.
Should you stop?

Let's start with the naive version of the question: line up each player's games and look at the one right after a streak. In the 2.67 million games here, the game after six or more straight losses scores 22 points per 100 games less than the game after six or more straight wins (43% against 65%). That is the kind of number articles about tilt usually report: one analysis of 840,000 Lichess games finds a 19-point swing in win rate between five straight losses and five straight wins, and a study of a million games finds that even the result from seven games back still predicts the current one. Almost none of it survives four checks:

• Opponents. During a winning streak of six or more, the opponents were on average 141 rating points weaker than the player; during a losing streak of six or more, 61 points stronger. The typical (median) streak game is much closer (48 points weaker, 17 points stronger): the averages are pulled by a tail of lopsided pairings, arenas, fresh accounts, the thin top of the pool, a friend who keeps accepting, which produce many long streaks. The next opponent comes from the same place: 175 points weaker on average after the wins (median 63), 61 points stronger after the losses (median 16). Keep only next games against an opponent within 100 points (that drops 47% of the games after a winning streak and 28% after a losing streak) and the 22-point swing shrinks to 6.
• The calendar. Seven in ten streaks of six or more span more than one sitting (a break of an hour or more starts a new sitting), and four in ten more than a day. A bad mood can last that long, but over days a streak also picks up everything else that drifts: your form, your rating, the pool you are being paired from. So this study asks the narrower question the "stop after three losses" advice is about: what a streak does to your next game while you are still at the board. Keep only streaks that happened in one sitting.
• Accounts. Players far from their usual level (second accounts, rating resets, three years of drift) are over-represented in long streaks. Drop each player's first 100 games in the record and any game played more than 150 points from their long-run rating.
• The rematch. About one next game in ten is against someone you already played during the streak, usually the same opponent again and again. The matchup that produced the streak is still on the board, and the opponent is in a state of their own; inside a rematch the streak effect is about twice as large. Keep only next games against a fresh opponent; the rematch gets its own treatment in section 6.

What survives all four is a drop of about 2 points per 100 games below normal in the game after six or more losses, the sliver labelled "maybe tilt" in the chart. After six or more wins its mirror image, +1.7, what sports science calls the hot hand. "Normal" is what players at your rating score against that opponent. It sits a little above 50% because the players in this database score 50.8% on these games on average (they are slightly more often the higher-rated side), and it tilts with the streak: streaks come from lopsided stretches, and the next opponent usually comes from the same stretch, after a losing streak rated a few points above you, after a winning streak a few points below (the dashed line in the chart). Both numbers are small. This story calls the shortfall after losses tilt for short, but it is a shortfall against a benchmark, not a reading of anyone's mood: fatigue, distraction and a bad connection all feed into it, and only players who kept playing are counted.

Score in the next game (points per 100 games), by the streak that just ended

Points above or below normal, one streak length at a time. A single loss costs almost nothing. Two to four losses cost half a point to just under one point per 100 games, with no clear order among them (the intervals overlap), five cost 1.6, six or more cost 2.0. The hot hand mirrors it, at up to 1.7 after six or more wins. To put those numbers in perspective, after three straight losses you still win the next game as often as you lose it, and even after six losses in a row the chance of losing the next one is about 50%, against 47% after a single loss. Comparing each player only with their own games, at the same clock setting and rating gap, gives the same picture (technical report, section 2d). If anything the shortfall is understated: six losses pull your rating below your real strength, so against an equal-rated opponent you should then score slightly above normal, not below.

Points above or below normal in the next game, per 100 games

After three or more losses in a row, rapid players fall 1.8 points per 100 games below normal, bullet and blitz players under one point. By rating, the shortfall is largest at the bottom, about 2 points at 800, and under a point from 1200 all the way to 2400. The hot hand is about one point everywhere. These are descriptions of groups, not of causes: rapid players and bullet players are different people playing at different clocks, and neighbouring groups are too close to tell apart.

Points above or below normal after 3+ losses and after 3+ wins, per 100 games

The obvious remedy is to walk away, and the data on what happens next is awkward: players who start their next game within a minute of a second straight loss score about normal, while players who pause for 10–30 minutes score 2.4 points per 100 games below it. That is not a verdict on breaks. The player who steps away after two losses is often the one who is more upset, more tired or more distracted, and a ten-minute break is more likely to be a phone call than a meditation. Players coming off wins show a similar, smaller dip after a pause, and the gap between the two groups closes after half an hour or so, but mostly because the winners come down too, not because the losers recover. From half an hour on, both lines sit about two points below normal whatever the last result, for a reason that has nothing to do with the streak: coming back cold. In this study a break of an hour or more starts a new session, and the first game of a session carries a cold-start penalty of its own (bullet −2.1, blitz −0.7, rapid −0.1 points per 100 games); the 30–60 minute cells show that the warm-up cost already sets in before the hour is up. "Normal" here is the level of games played mid-session, so a game after a long pause sits below it whether the last game was a win or a loss. How long a streak's effect lasts for a given player, and what a break does to it, the data cannot say.

Two more things the data do say. Most players keep going: after a loss, seven in ten games are followed by another within the hour, and after six or more losses nearly nine in ten, mostly because long same-sitting streaks happen deep in long sessions. And bullet has a twist: playing straight on after two losses costs nothing, a 3–10 minute pause is followed by −2.9, and the first game of any bullet session runs −2.1. A cold start costs a bullet player more than a losing streak ever does.

Points per 100 games above or below normal, after 2+ losses or 2+ wins, by the break before the next game

From half an hour on, both lines show the warm-up cost of coming back cold, not an after-effect of the streak; breaks of an hour or more start a new session.

This section sets the streak aside. It takes every loss, whatever came before it, and asks what the next game looks like depending on how that one loss went. Folklore says the blown win hurts most. Split losses by what they looked like and the data cannot find it. A loss in which you had entered the endgame winning (a Stockfish evaluation of +2 or better, and still lost) is followed by a game at or above normal (+0.4). Two thirds of those are flags in a won endgame; the blown win by resignation, the kind that really stings, is rarer, too rare to pin down exactly, but nothing large is hiding there. A long fight of more than 30 moves: no effect at all (+0.1). What is followed by a worse game is the quick collapse. A loss in under eleven moves is followed by −2.1, a game that never reached an endgame by −1.2, and a loss by disconnection by −3.7 (though that last number is mostly the connection, not the player; see below). Resigning (−0.6) is worse than getting checkmated (−0.2) or flagged (−0.3).

A plausible reading is not that quick losses cause tilt but that they reveal something: a ten-move loss is often a distracted, pre-moving, phone-in-hand game, and whatever it was is still there in the next one. The disconnect cell shows this directly. After a loss by disconnection, one next game in twelve is itself abandoned, four times the usual rate, and those games score 30 points per 100 below normal, four in five lost: the connection was still flaky. Keep only next games that were played to a result and the cell drops to −1.1, no worse than any other loss. The short losses survive the same cut (−2.1 either way), and they also hold when each player is measured against their own usual performance rather than their rating (−1.6), so the cell is not a few chronic under-performers. Whether a distracted evening carries into the next game, as a bad connection does, the data cannot separate from tilt. One caveat: a game thrown away before the endgame, a hung queen at move 20, lands in the "never reached an endgame" cell with every other short loss, so this cut cannot tell that kind of blown win from a collapse. What it can say is that among players who kept playing, the game you fought hard and lost is not followed by a worse one. The wins mirror this: long wins are followed by better games than short ones (+0.9 after more than 30 moves, +0.2 under eleven), and a comeback win from an endgame entered losing (Stockfish −2 or worse) by +1.0; both are in the technical report, section 5.

Points above or below normal in the next game, by what the loss looked like, per 100 games

If the score barely moves, where does the feeling go? Into behaviour. Like section 5, this one sets the streak aside: it compares what players do after any loss with what they do after any win, whatever came before. After a loss, players:

• Speed up. Time per move in the opening and early middlegame (moves 3 to 20) falls 1.5 to 4% after a loss compared with after a win, comparing each player with themselves. It is not a few people driving the average: two thirds to three quarters of players are individually faster after a loss.
• Blunder a little more, eventually, maybe. In rapid games with an engine evaluation (a Lichess analysis requested by either player, or Stockfish run by the study), 5.8 blunders per 100 moves after a loss and 5.7 after a win; after three or more losses, 6.3. The opponent's rate rises in those same games too, by about half as much, so part of this may be the kind of game rather than the player. On the clock the change shows after one loss; on the board, if at all, after three.
• Quit, or rush. A bullet session ends after 22% of losses and 17% of wins. Rapid players who carry on do it faster after a loss: of those who play again within the hour, half start within a minute after a loss and four in ten after a win.
• Take revenge. Every analysis so far excluded rematches and counted only fresh opponents; here the rematch comes back. One next game in eleven is a rematch against the opponent who just won. Those games run 1.8 points per 100 games below normal, against 0.5 for the next game against someone new, so the rematch itself costs about 1.3 points. How much of that is the matchup and how much is you, the data cannot split: the player who just beat you is, on the evidence of that game, a little better than their rating, and the rematch against someone you just beat shows a matching bonus, but both players choose the rematch for reasons of their own. The penalty grows with the time control: −1.2 in bullet, −1.7 in blitz, −3.0 in rapid, −4.3 in classical.

The revenge rematch, and who quits on a loss

• Your stats overstate tilt. After three straight losses your next game runs 0.5 points per 100 games below normal, about one draw turned into a loss per 100 games; after six or more, two wins turned into losses. Your raw stats show a swing several times bigger, and that part is lopsided pairings and the calendar, not you.
• The stop-loss rule is untested, not disproven. Games played right after two losses in one sitting cost you less than one point per thousand games in total, and whether stopping would recover any of that, no game history can show, because the games that stopped players would have played do not exist. What the data do show: in bullet, a cold start (2.1 points for the first game of a sitting) is bigger than any losing streak.
• Think twice about the revenge rematch. The immediate rematch against the player who just beat you runs about 1.3 points below an ordinary post-loss game, more than four losses in a row, and the price rises with the time control.
• Do not fear the long loss. Losing a game you fought for, even from a won endgame, is not followed by a worse game on average. The losses that are: ten-move collapses, which mostly tell you that you were not really playing in the first place. A disconnect only predicts another disconnect.
• Watch the clock first. After a loss you play 1.5 to 4% faster and are more likely to quit or to rush into the next game; the blunder rate only moves, if at all, after three losses in a row. If you notice yourself pre-moving through the opening, that is what the data can see.