Why Three Models Instead of One
A single machine learning model captures the world through its own structural assumptions. Random Forest excels at finding which input features matter most. LSTM neural networks are built to find sequential patterns across time. XGBoost is optimized for tabular data where interactions between many features drive the output. No single architecture is best at all three tasks simultaneously — so we run all three and combine their outputs.
What LSTM Contributes
Long Short-Term Memory (LSTM) networks are a type of recurrent neural network designed to learn from sequences. For lottery data, the "sequence" is the ordered history of draw results. The LSTM learns patterns like: "number 28 tends to appear within 3 draws of number 14" or "position 5 numbers show a mean-reversion tendency after gaps longer than 25 draws." These are weak statistical signals — not predictive in the usual sense, but real patterns in historical data that the model can weight.
What Random Forest Contributes
Random Forest trains hundreds of decision trees on different subsets of our 115 engineered features (frequency, gap, positional averages, parity ratios, range statistics). Each tree votes on which numbers are most "consistent with current statistical conditions." The forest's output is a probability-weighted ranking of candidate numbers, robust to overfitting because it averages across many trees.
What XGBoost Contributes
XGBoost (Extreme Gradient Boosting) iteratively corrects the errors of previous trees, making it especially good at capturing non-linear feature interactions. In practice, this helps identify combinations of conditions — for example, "when gap is above 20 AND positional frequency in position 3 is below 0.04 AND recent sum average is above 155" — that no single feature reveals alone.
The Ensemble Layer
A meta-learner (another lightweight model) takes the outputs of all three base models as its inputs and learns how to best weight them for the specific game and time window. This approach, called stacking, consistently outperforms any individual model on held-out validation sets in our testing.
The Honest Limitation
Lottery draws are designed to be statistically independent events — each ball is drawn from a freshly loaded machine with no memory of past draws. Any patterns an ML model finds in historical data are artifacts of finite sample size, not causal laws. Our models find the strongest available statistical signals in historical draws, but they cannot change the fundamental randomness of each new drawing. Use these tools as one informed input among many — not as a guarantee.