Footprint ML

Four Machine Learning Models

1. LSTM (Long Short-Term Memory): A specialized neural network designed to capture long-term dependencies in sequential data by using a gating mechanism to selectively remeber or forget information over time.
2. RNN (Recurrent Neural Network): A foundational neural network alghoritm that processes sequential inputs by passing a hidden state forward.
3. MLP (Multi-Layer Perceptron): A feedforward neural network that stacks layers of artificial neurons to map complex, non linear relantionship between inputs and outputs.
4. GBDT (Gradient Boosted Decision Trees): A ensemble machine learning alghoritm that builds simple decision trees sequentially with each new tree correcting the errors and residual made by the previous ones. 

1. Base Learn trains a model using the configured parameters; training runs across multiple epochs and walk-forward sets. 
2. Genetic Learn uses an evolutionary alghoritm to evaluate candidate model configurations. Candidate fitness is determined from validated best-checkpoint results across the walk-forward sets. This allows the system to explore different model configurations and retain stronger candidates based on validation performance.
3. Deep Learn evaluates a structured grid of configured model and training parameters. Candidate models are compared using validation results, and the strongest validated configuration is retained.

Two Prediction Modes

1. Classification mode estimates the direction of the future price move over the configured prediction horizon. The model produces a signed output bounded by its configured activation function (-1/+1). Performance is evaluated using directional accuracy.
2. Regression mode estimates the future return as a percentage. Regression candidates are evaluated primarily using validation error, with a custom score based on RMSE and  Directional-match statistics. 

Tools

1.  Features Extraction: when enabled, the system evaluates candidate input sets and trains the final model using the selected features returning the best result.                                                                                                                                                                                                                                                                                              -Genetic Extraction searches for effective feature subsets using an evolutionary optimization process; this allows the system to test combinations of features and identify stronger input sets.                                                                                                                                                                                                                                       -Sequential Extraction evaluates ranked feature prefixes to identify an effective subset size.                                                                                                                                                                                                                                                                                                                                                                                                           -Top Extraction ranks features by importance and selects the configured number of top-ranked inputs.

2. Walk-Forward Training and Validation: Historical data is partitioned chronologically into walk-forward sets; each set contains its own training window and validation window; for every walk-forward set, the best-performing epoch is selected based on validation evidence. Candidate configurations are ranked using their selected checkpoints, not simply the weights from the final epoch; this approach evaluates model behavior across multiple historical windows instead of relying on a single training-validation split.
3. Early-Stop / Deep-Patience: Deep patience stops candidate training after the configured number of consecutive epochs without improvement in its validation selection score; early stop stops candidate training when the configured validation metric reaches or exceeds the target threshold.
4. Multi-Timeframe Features Architecture: The indicator allows the feature timeframe to be configured separately from the chart timeframe; completed source-timeframe bars are aligned with chart-timeframe observations so the model can use multi-timeframe context when predicting future candles on the chart timeframe. This structure helps maintain chronological consistency while allowing the model to access information from different timeframe resolutions.
5. Paired Symbol Features: An optional paired symbol can provide cross-market context during feature construction and training.
6. Ridge Calibration: Ridge Calibration is an optional post-training refinement stage applied to the model’s raw predictions; it uses L2-regularized linear regression to learn a stable mapping from the base model output to the final calibrated prediction. The Ridge penalty constrains the calibration coefficients, reducing sensitivity to noisy or unstable model outputs and helping avoid overfitting to small variations in raw scores.
   
7. ML Graph: The visualization graph displays relationships between input features, training epochs, candidate models, validation sets, and calibration stages; connected nodes and highlighted paths illustrate how candidate model lineages evolve and lead to the selected model neural network. 

Note: MT5 provides native historical tick-by-tick data, allowing footprint rows to show reported real traded volume executed at each price level when the broker supplies real-volume data.