Training & Data

Guardrails & Training Reference

Understand the protections NeuraLib applies to time-series data, then choose activations, objectives, metrics, optimizers, and schedules that match your task.

Updated July 12, 2026Publisher Alien_AlgorithmsPine Script v6

Built-in data guardrails

Invalid rows are rejected: wrong-width rows and rows containing na never enter a dataset.
Shape checks protect calls: forward, training, backward, and evaluation paths validate tensors before graph execution.
Train and validation remain separated: holdout extraction is chronological.
Scaler leakage is controlled: validation transforms use the training-side profile.
Rolling windows preserve time order: chronological reads remain correct even after ring-buffer wrapping.

Guardrails catch mechanics, not research assumptions

NeuraLib can reject a malformed row, but it cannot know that your feature definition contains future information or that your chosen target is economically meaningless.

Activations

nl.ActivationKind includes Linear, ReLU, Leaky ReLU, ELU, GELU Approx, Sigmoid, Tanh, Softplus, Swish, and Softmax.

Family	Typical role
Linear	Regression output or raw logits.
ReLU / Leaky ReLU / ELU	Efficient hidden-layer nonlinearities.
GELU Approx / Swish	Smooth nonlinearities for deeper or attention-style models.
Sigmoid	Binary probability output when the loss expects probabilities.
Tanh	Bounded signed hidden state or output.
Softmax	Multi-class probabilities; prefer a from-logits loss during training when available.

Losses

Loss	Typical use
MSE, MAE, Huber, LogCosh	General regression. Huber and LogCosh reduce sensitivity to large outliers.
Binary Cross Entropy	Two-class probability output.
Binary Cross Entropy From Logits	Numerically stable two-class raw-logit training.
Categorical Cross Entropy	Multi-class probability output.
Softmax Cross Entropy From Logits	Numerically stable multi-class raw-logit training.
Directional	Finance-oriented penalty for incorrect sign.
Quantile	A chosen conditional quantile instead of the mean.
Multi-Horizon Weighted	Multiple forecast horizons with explicit weights.
Sharpe	Sharpe-style risk-adjusted objective.

Metrics describe behavior without driving gradients. Available kinds include MAE, RMSE, Directional Accuracy, Binary Accuracy, Binary Accuracy From Logits, Categorical Accuracy, and Cosine Similarity.

Optimizers and schedules

Optimizers

sgd(), momentum(), rmsprop(), adam(), and adamW() cover simple through adaptive update rules.

Schedules

constantSchedule(), stepDecay(), cosineDecay(), and warmupCosine() control the current rate, which can be inspected with currentRate().

Choosing an objective

Match output activation and loss expectations: logits with from-logits losses, probabilities with probability losses.
Use robust regression losses when rare market moves dominate MSE.
Keep a human-readable metric alongside the training objective.
Compare against a simple baseline; lower loss is not automatically useful trading information.