Random Baseline: The Only Number That Matters in Lottery AI

Here is the only question worth asking about any lottery prediction model: does it beat a coin flip? Not on the data it trained on — on draws it never saw. That distinction is everything. A model that memorizes history looks brilliant in hindsight and useless in practice. Jackpot Teller's evaluation system exists to answer this question without flinching. We run each trained model against a simulated random picker on the same held-out draws, then report a hard verdict. In the latest evaluation snapshot, 13 of 60 models cleared that bar. The other 47 did not. We publish both numbers. That is the whole philosophy in two sentences: measure against random, tell the truth.

What 'Beats Random' Actually Means

The phrase beats random sounds simple. It is not, once you care about getting it right.

When a model finishes training, it gets evaluated on a held-out set of recent draws — draws deliberately withheld during training so the model has never seen them. A simulated random picker then plays those exact same draws. If the model's hit rate on that holdout exceeds the random picker's hit rate on the same draws, it earns the Beats Random verdict. If it doesn't, it's labeled Below Random.

The held-out set is typically around 20 draws. That is a small window, which is why we don't stop at a point estimate. Hit-rate uncertainty is reported using a 95% Wilson confidence interval — a method that handles small samples and extreme proportions more honestly than a naive percentage would. The interval tells you how wide the uncertainty band is around the observed hit rate, which matters enormously when the holdout is modest in size.

• Holdout isolation: The model never touches evaluation draws during training.
• Same draws, same conditions: Random picker and model face identical draw sets.
• Confidence interval: Uncertainty is quantified, not hidden.

The median holdout hit rate across all evaluated models sits at 70.0%. That sounds high until you realize the random baseline is calculated on the same draws — so what matters is not the absolute rate, but whether the model's rate clears the random picker's rate on those specific draws. A 70% hit rate means nothing in isolation. Context is everything.

Why 47 of 60 Models Fail — and Why We Say So

In the latest evaluation snapshot, 47 of 60 models were labeled Below Random. That is a 78% failure rate, and we are not embarrassed by it. We are proud of it, because it means the evaluation is working.

Think about what the alternative looks like. A system that only publishes models when they look good is a system that has quietly discarded its failures. You never see the graveyard. Jackpot Teller's admin dashboard shows the graveyard. Every model, every verdict, every confidence interval — including the ones that lost to a random number generator.

The models that do pass are built on real signal. Our database holds 568,593 draws spanning draw history back to 1980, across 355 games. The single largest model in the system trained on 16,549 draws for one game — that is decades of a single draw's history compressed into one training corpus. Sequence models only train where history is long enough to support them; games with thin records don't get forced into architectures that require depth they can't provide.

The ensemble approach — random forest and gradient boosting (XGBoost) models per game, plus frequent-itemset mining — means no single algorithm gets to dominate. Each game gets the architecture its history supports. Models retrain daily after new draws are ingested; games with no new draws are simply skipped rather than retrained on stale data.

If you want to explore what a model-backed pick actually looks like when one does clear the random baseline, our AI/ML predictions page surfaces only the games where a current model holds a Beats Random verdict. No verdict, no pick — that is the rule.

Pattern study across this volume of history is analytically interesting. Whether any edge persists into future draws is a separate, harder question — one the holdout evaluation is designed to probe honestly, without guaranteeing anything.

What This Means When You Read a Verdict

When you see a Beats Random label on a game in Jackpot Teller, you now know exactly what it took to earn it: a model trained on historical draws, evaluated cold on draws it never saw, outperforming a simulated random picker on those same draws, with uncertainty quantified at a 95% confidence level. That is not a marketing claim. It is a measurement with a methodology you can interrogate.

When you see Below Random, that is equally informative. It tells you the model found no statistically meaningful edge on recent holdout draws — and that using its picks would be, at best, equivalent to a quick pick. There is no shame in a quick pick; randomness is a perfectly coherent strategy. The point is that you should know which one you are using.

The 13 models that currently beat random represent a minority of the 60 evaluated — and that minority changes as models retrain daily and draw history grows. A game that fails today may pass next week. A game that passes today may slip below the line after its next evaluation. The dashboard reflects the current snapshot, not a permanent ranking.

For the data-curious reader, that dynamism is the most interesting part. This is not a static leaderboard. It is a live experiment, run every day, against the only benchmark that cannot be gamed: genuine randomness.