Monte Carlo Blackjack Simulator

Simulation engineering, experiment design, and statistical honesty — Python, pytest, reproducible by construction.

Report (PDF)Code

Every result in the blackjack series rests on one piece of infrastructure: a simulator whose numbers can be trusted. Blackjack is one of the rare domains with published ground truth — house edge, dealer bust rate, blackjack frequency are known to decimal places — so the engine had to validate itself against the literature first. If those numbers didn’t match, nothing downstream would mean anything.

The design

Any playing strategy — from a hard-coded rule set to a neural network — plugs in through one method, decide(state) → action, against an immutable game state. Betting strategies, counting systems (Hi-Lo, KO, Omega II), and casino rule sets swap independently. This is what later made the RL work honest: every learned agent plugs in as a strategy object with no engine changes, directly comparable to every result here — same instrument, same terms. Runs are reproducible by construction: fixed seeds, independent random streams per strategy, and one command that regenerates all ~80 million hands.

What the instrument showed

  • Strategy is worth ~45 points of house edge. Basic strategy plays at a 0.45% house edge; random play at 45.8%.
  • Avoiding busts is not the goal. The strategy with the lowest bust rate (12.4%, vs. basic strategy’s 15.8%) loses money twelve times faster. Maximizing expected value is the goal; busting less is just one term in it.
  • Betting systems move risk, not expectation. Flat betting and Martingale lose the same money per 1,000 hands (−$45 vs. −$48) — but Martingale ruins 40.1% of sessions to flat’s 1.0%. Expected value and risk of ruin are independent levers; profit is not survival.
  • Counting is the only genuine edge — roughly +1.3% per dollar wagered on a single deck. The information is the edge, not the betting pattern.

Getting there included a correctness audit that caught real, silent bugs — split hands scoring only the first hand, split aces illegally re-hitting for a phantom ~0.28% of edge. The habit of auditing the instrument before believing its output became the method for everything that followed.