A mathematical function that takes inputs (features) and outputs predictions (here: P(win)). Our model is "trained" to make better predictions by adjusting internal numbers (weights) based on past wins and losses.

A single numeric input the model uses to make predictions. We use 22 features: RSI, ATR%, RVOL, MA distance, IV, regime, etc.

A number that says how much a feature matters to the prediction. Positive weight = feature pushes toward "win"; negative = toward "loss." Training adjusts these weights to minimize errors.

The "correct answer" we're trying to predict. Here labels are binary: 1 = hit (won), 0 = miss (lost). Each finding becomes a labeled training row after the brain rates it 30 min later.

The simplest classification model. Takes a weighted sum of features, squashes it through a sigmoid function to get a probability between 0 and 1. Fast, interpretable, hard to overfit. Our primary model.

Multi-layer perceptron. Stacks logistic regressions with non-linear activations (ReLU) so it can learn feature interactions. More powerful but slower, harder to interpret, more prone to overfitting.

The algorithm that updates weights one sample at a time. For each training row: predict, compare to actual, nudge weights to reduce the error. Repeat thousands of times.

How big each weight update is. Too small: model trains slowly. Too big: oscillates / never converges. Sweet spot is typically 0.01-0.1. See LR Tuner.

One complete pass through all the training data. We typically do 3-5 epochs in a full retrain. More = better fit but risk of overfitting.

A number measuring how wrong the model's predictions were. Lower = better. We use binary cross-entropy which heavily penalizes confident-but-wrong predictions.

When the model memorizes the training data instead of learning generalizable rules. Shows up as high in-sample accuracy but low out-of-sample. Mitigated by: simpler model, more data, regularization.

(Correct predictions) / (Total predictions). Random guessing = 50%. Anything above is signal. Above 55% is a meaningful trading edge.

When the model says "WIN," how often is it actually a win? Of 100 buy signals, if 65 are wins, precision = 65%.

Of all the actual wins, how many did the model catch? If 100 setups are real wins but model only flagged 70, recall = 70%.

Harmonic mean of precision and recall. Single number balancing both. Above 0.55 is solid for trading models.

A 2×2 table: True Positives (predicted win, was win), False Positives (predicted win, was loss), True Negatives, False Negatives. Tells you what KIND of errors the model makes. See Prediction Replay.

Robust accuracy estimation. Split data into K folds. For each fold: train on the other K-1, test on this one. Average results. Tells you how the model will perform on unseen data.

Training one sample at a time as new data arrives. Our brain does this on every newly-rated outcome — model keeps improving as you use the site.

Designing what to feed the model. Bad features = bad predictions. Good features = strong predictions even with simple models. See all features documented.

Running multiple models and combining their predictions. Reduces variance, often improves accuracy. See Ensemble A/B + Model Compare.

A method for explaining individual predictions: shows which features pushed the prediction up (toward win) or down (toward loss). Tells you the WHY behind any model output. See Model Explain.

The model's output — a number 0-1 representing P(win). 0.7 = model thinks 70% chance this works. Not certainty — calibrated probability. See Model Confidence.

A fresh model has no data — predictions are random. Takes 50-100 real outcomes before it's calibrated. Use Model Seed to bootstrap w/ synthetic but realistic data.

Initialize a fresh model with pre-existing knowledge instead of zero weights. Our Model Seed feature does this by generating realistic synthetic data based on historical setup hit-rates.

P > 0.7 = high confidence (model lifts conviction). 0.45-0.55 = noise (model has no opinion). P < 0.3 = model says fade. Always interpret in context of overall strategy.

Every snapshot of weights is a version. Old versions stay archived (up to 50). Promote any past version to current with one click if a training run goes wrong. See Model Versions.