AFML Bet Sizing with Confidence, Concurrency, and Exposure Controls
Summary
This article presents four bet-sizing methods for turning trading signals into positions: probability-based sizing, dynamic sizing from forecast-price divergence, budget-constrained sizing, and reserve sizing fitted to concurrent position imbalance. The probability method maps classifier confidence relative to a no-skill baseline into a bounded signed signal. It can average overlapping active bets and discretize changes to limit turnover. The other methods address forecast-price inputs or directional signals when classifier probabilities are unavailable.
The article explains why fixed-fraction rules can ignore confidence, payoff asymmetry, overlapping positions, and trading costs. It recommends choosing a sizing curve to fit the expected relationship between forecast divergence and value, and calibrating its aggressiveness from an explicit target. The methods do not themselves correct poorly calibrated probabilities or fully account for payoff ratios; the article points to a separate Kelly-based treatment for that. Although it describes a runnable implementation and gives illustrative examples, it does not present an independent performance evaluation establishing that these sizing rules improve live results.
Key ideas
- Fixed position fractions ignore differences in signal confidence and payoff structure.
- Overlapping labels can cause aggregate exposure to rise sharply unless active positions are accounted for.
- Probability-based sizing maps confidence relative to a base rate into a bounded signal.
- Discretizing position updates can reduce turnover when small changes are unlikely to cover trading costs.
- Dynamic, budget, and reserve methods address different forecast inputs and concurrency constraints.
- Sizing quality depends on calibrated probabilities and realistic assumptions about payoffs and execution costs.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.