MAE
Historical mean absolute error. It shows how far, on average, the model is from the observed price. Lower is better.
PredictBroker, explained for humans.
LSTM
Neural network designed to learn patterns in sequences and time series.
Ensemble
Combines several models and gives more weight to those with lower historical error.
Confidence
Indicative 0–100 score: 60% skill (model historical error vs. the volatility-implied move) and 40% agreement between models on direction. It is not a guaranteed probability of profit.
Maximum Drawdown
Largest fall in capital from a peak to a later trough during the simulation.
What‑If
Historical simulation based on predictions saved before the outcome was known. It does not represent guaranteed profit.
Direction accuracy
Share of verified predictions that got the direction (up/down) right. 50% is a coin flip.
Backtest
Walk-forward test: the model is trained only on data before each date and compared with the real outcome. It is generated after the fact, so it never counts towards the live Track Record, Success Stories or What-If.
Calibration
Compares stated confidence with realised accuracy. If confidence is useful, bars should rise to the right.
Signal return
Average return per prediction if you had followed the predicted direction (long if up, short if down), before costs.