Seven wins in ten games: does that prove you improved?

1. How much noise is in ten games?

A single result is driven by many random factors: the deal, the draw, opponents’ choices, seat and rules. Seven wins out of ten may come from skill — or from good tiles and generous opponents.

The reverse is also true. Seven losses do not mean you suddenly got worse. The core problem with a small sample is that it is not repeatable: change the day or the table and the result can flip.

2. Result metrics vs process metrics

Result metrics are win rate, deal-in rate, average placement and average hand value. They matter, but they are noisy and situation-dependent.

Process metrics are closer to what you actually did: did you miss visible information, were your danger-tile calls consistent, did you have a reason for switching between attack and defence, could you explain the alternative in review?

Pick two process metrics per week. “Deal-in rate after riichi” and “pointless riichi chases in the final round” are enough. Too many metrics and the logging becomes the work.

3. Use a rolling window

Do not stare at lifetime win rate. Look at the last 50 and 100 games. Plot a point every 20 games and draw a simple trend line. A rising trend says more than one 7–3 run.

If you only have a dozen games, write “not enough data”. That is not pessimism; it keeps noise from setting your training plan.

4. A log you can actually keep

After each session, write three lines: one key decision, one result, one alternative. Do not try to record everything.

Every 20 games, summarise: which error keeps repeating? Did you fail to see information, or see it and judge it wrong? The first is a process fix; the second is a training fix.

Every 100 games, look back again. It becomes much easier to separate a short-term swing from a stable habit.

5. When should you change your style?

When the same process metric gets worse across several windows, and your review finds the same cause. Changing style just because you “keep losing” usually turns variance into a feedback loop.

Change one controllable action at a time, then observe at least 50 games. One variable at a time is the only way to know what caused the change.