Optimal Blackjack Betting Strategies Through Dynamic Programming and Expected Utility Theory

TL;DR

This work formulates blackjack as a pair of coupled Markov decision processes to optimize round-by-round decisions and stake sizing under risk-sensitive utility. By constructing origin-deck, player/dealer sequences, and transition dynamics, it derives both a restricted, exact solution for a simplified round policy and approximations for handling the full Split option, deploying dynamic programming and mass-vector techniques. The results show that a semi-optimal round policy offers modest gains over Basic Strategy with higher volatility, while deck-composition-based betting policies (CRRA/CARA) yield favorable return distributions and better risk-adjusted metrics in simulations. The study demonstrates the feasibility of near-optimal betting strategies under specific blackjack rules, with implications for finance-like decision-making and potential extensions to more complex game configurations and multiplayer settings.

Abstract

This study presents a rigorous mathematical approach to the optimization of round and betting policies in Blackjack, using Markov Decision Processes (MDP) and Expected Utility Theory. The analysis considers a direct confrontation between a player and the dealer, simplifying the dynamics of the game. The objective is to develop optimal strategies that maximize expected utility for risk profiles defined by constant (CRRA) and absolute (CARA) aversion utility functions. Dynamic programming algorithms are implemented to estimate optimal gambling and betting policies with different levels of complexity. The evaluation is performed through simulations, analyzing histograms of final returns. The results indicate that the advantage of applying optimized round policies over the "basic strategy" is slight, highlighting the efficiency of the last one. In addition, betting strategies based on the exact composition of the deck slightly outperform the Hi-Lo counting system, showing its effectiveness. The optimized strategies include versions suitable for mental use in physical environments and more complex ones requiring computational processing. Although the computed strategies approximate the theoretical optimal performance, this study is limited to a specific configuration of rules. As a future challenge, it is proposed to explore strategies under other game configurations, considering additional players or deeper penetration of the deck, which could pose new technical challenges.

Figures (9)

• Figure 1: Comparison of the estimated expected return under the semi-optimal policy and Basic Strategy. The first subfigure analyzes the 50,000 sampled decks, while the second focuses on favorable decks.
• Figure 2: Analysis of the estimated standard deviation under the semi-optimal round policy and Basic Strategy. The first subfigure analyzes the 50,000 sampled decks, while the second focuses on the convenient decks.
• Figure 3: Histogram of relative frequencies of the true count for the 50,000 sampled decks.
• Figure 4: Comparison between the estimated expected return and the true count under semi-optimal and Basic Strategy round policies. The first subfigure shows the relationship across the entire true count range, while the second focuses on values associated with a positive expected return.
• Figure 7: Heatmaps of the estimated optimal auxiliary policy as a function of wealth and the index of the auxiliary mass vector, for $n = 50$ rounds and under three risk levels $\beta$: 0.15, 0.25, and 0.35.