Betting Against the Noise
Adversarial self-falsification, positive and negative controls, and multi-axis risk evaluation — PyTorch, Kelly / log-growth analysis.
You have a genuine edge on a bet — how much should you wager? There’s a formula for the optimal amount, known since 1956: the Kelly criterion. This study asks whether a model could learn that sizing rule on its own — from nothing but its own wins and losses — instead of being handed the formula. Blackjack is the cleanest possible testbed: card counting is a rare case where the edge can be measured precisely (here, over 20 million hands — break-even lands at true count +0.76, right where the folklore says) and the optimal bet ladder is cleanly derivable.
The learned agent never found it. Across longer training, bigger batches, and every standard fix, it collapsed to the same policy: bet the minimum, every time. It sometimes drifted through Kelly-shaped sizing curves — an orbit that visits and leaves — but head-to-head, those wanderings lost to betting flat.
Testing the failure like a success
A model that fails isn’t a finding; the reason is. First, a tempting story: the network’s internal representation, plotted, split into beautiful clean clusters — by bankroll, not by the edge. Obvious conclusion: it had learned to bet on how much money it had. Then the story was tested instead of believed. Remove the bankroll input entirely: no change. Retrain across every bankroll: no change. The beautiful explanation was wrong, twice — nothing wealth-shaped was load-bearing.
That left two suspects: the machinery or the signal. An oracle control settled it — the same network fed a denoised reward learns the Kelly ladder immediately and holds it. The wall wasn’t the model. It was the information.
The information is brutal
Three counts. The prize is tiny: over a 1,000-hand session, perfect sizing beats flat betting by about 1% of growth. The signal is buried: it must be estimated through per-hand noise roughly fifty times larger than the strongest edge. And the carriers are rare: the counts holding most of the edge appear in under 1% of hands. The model isn’t under-powered — it’s being asked to hear a whisper, in a storm, that almost never speaks.
The 1956 formula wins by using the same information differently: measure the edge once, offline, then derive the bet. “The edge is real” and “the edge is learnable from experience” are different claims, and the gap between them is measurable. When the optimum is derivable, derive it — and thin, rare, sub-noise edges are exactly what outcome data looks like almost everywhere people point learning algorithms at money.