Exploitability and model limits
Measure how much a best response can gain in a specified zero-sum game.
12 minutes · advanced · Original Poker4Quant lessonWhat you will learn
Compute a one-sided best-response gap and state what its units and reference value mean.
Prerequisites: Mixed strategies · A bluff-catch toy game
A gap needs a reference
Suppose a zero-sum game has row equilibrium value v. For a proposed row strategy, compute its worst expected payoff over column responses. The difference v minus that worst value is a one-sided shortfall from row’s equilibrium guarantee.
Terminology for exploitability and combined two-player deviation gaps varies. State exactly which quantity you report. A raw number without the game, payoff units and baseline is not a transferable measure of a real poker player’s skill.
Model error can dominate a precise solution
In matching pennies, row’s H-probability p gives worst value −|2p−1|. At p=0.70 the worst payoff is −0.40, versus equilibrium value zero, so its one-sided shortfall is 0.40 per play.
This clean result relies on known simultaneous actions and exact payoffs. If the real interaction is sequential, if observations leak, or if utility is not the listed zero-sum payoff, the model’s guarantee may not apply. Solving the wrong model exactly does not repair those differences.
Worked example
A matching-pennies row player chooses H 60%. Find its worst value and one-sided shortfall from v=0.
- Against column H, row earns 0.60−0.40=0.20.
- Against column T, row earns −0.60+0.40=−0.20.
- A minimizing opponent chooses T, so worst value is −0.20.
- Subtract that from the equilibrium value: 0−(−0.20)=0.20.
The one-sided shortfall is 0.20 payoff units per play. This is not a percentage of hands or a prediction against every actual opponent.
Common mistakes
- Reporting a gap without its baseline or units.
- Assuming a worst-case loss is the result against every opponent.
- Extending a toy-game guarantee to a richer real information structure.
Keep these ideas
- Compute the opponent’s actual best response in the stated game.
- Specify one-sided or combined deviation metrics.
- Check model fidelity before interpreting precision.
Further reading
- MIT OpenCourseWare — Mixed Strategy Nash Equilibrium, lecture 3 ↗
University lecture background on expected utility, mixed strategies and indifference. Poker4Quant’s toy payoffs and exercises are original; this link does not imply affiliation or endorsement. Checked 2026-10-21.
- NYU — Game Theory, Problem Set 1 (2003) ↗
University problem set on best responses, dominance and the limitations of weak-dominance elimination. Background reading; the exercises on this site use their own specified payoffs. Checked 2026-10-21.