Poker Decision Analytics

Analyzing 60,000+ poker hands to measure decision quality, variance and tail risk.

Problem

Judging a decision by its outcome is the most common mistake in any domain with uncertainty: a good decision can lose and a bad one can win, and with enough noise the two are indistinguishable in a single sample. What is needed is a dataset where the decision, the context it was made in and the measurable result are all recorded, so that decision quality can be separated from luck. Poker hand histories are one of the few public datasets with all three.

Approach

Treat the hand history as structured data rather than as a game. Every hand gives an action, a position, a betting round and a numerical outcome in big blinds, which is enough to estimate expected value conditioned on context, to measure dispersion rather than just the mean, and to locate where losses concentrate. The framework is deliberately about decisions under uncertainty, with poker supplying the data; the same shape of question applies wherever outcomes are noisy and context matters.

Architecture

The pipeline has three stages that are kept strictly separate: a parser that reads the plain text .phh files of the public Pluribus dataset and normalises them into two flat tables, one row per hand and one row per player-hand; a storage step that writes both as CSV and Parquet and commits the parsed result to the repository, so a fresh clone runs immediately without downloading ten thousand raw files; and an analysis layer built on pandas that computes expected value, variance, drawdown and tail risk, sliced by position and by street, emitting five figures. Splitting parsing from analysis is what makes the expensive step run once and the cheap step run as often as a question needs re-asking.

Measured results

The corpus is over 60,000 hands across 92 sessions. The aggregate result is -2,413 big blinds, which is expected and is discussed below rather than hidden. The risk profile is more informative than the mean: the average of the 5 percent worst outcomes is -14.24 big blinds, and the ratio of the average win to the average loss is 1.82, meaning the winning hands are substantially larger than the losing ones even though the total is negative. That asymmetry is the interesting finding, because it is the signature of a strategy that loses small and often while winning large and rarely, and it is invisible to anyone who only looks at the total.

Engineering decisions

Separating parsing from analysis and committing the parsed tables was the decision that made the project usable by anyone else: the ten thousand raw files are reproducible from a public source but nobody wants to download them to read a chart. Choosing Parquet alongside CSV keeps the tables both machine-efficient and human-inspectable. Reporting tail risk and dispersion rather than only the mean is the analytical decision that the whole project rests on, since a mean alone cannot distinguish a steady strategy from a volatile one with the same average.

Limitations

The dataset comes from the training of an AI agent in an exploratory phase, not from expert play, so the negative aggregate result is an expected property of the data and not a finding about poker. This is not a strategy tool and gives no advice about how to play. The analysis is observational and correlational: it can show that outcomes differ by position, not that position caused the difference, since the decisions were not randomised. Expected value is estimated from realised outcomes, which conflates decision quality with variance at any sample size, and the fixed six-player Pluribus setup means none of the numbers generalise to other formats.

At a glance