Best Practices & Limitations

Start small

• Begin with 4–8 hidden units and batches of 8–16.
• Prove that a dense baseline has signal before adding temporal or attention blocks.
• Add one architecture change at a time and compare validation behavior.
• Prefer stable feature definitions over a large feature count.

Protect causality

• Create features only from information available at their anchor bar.
• Delay target creation until the forecast horizon has resolved.
• Avoid centered indicators, negative offsets, or future-dependent normalization.
• Keep rolling windows chronological and align target offsets explicitly.

Validate honestly

• Use chronological holdouts rather than random shuffling.
• Fit scalers from training rows only.
• Track validation loss and a task-relevant metric.
• Compare with naive baselines such as last value, zero return, or majority class.
• Use early stopping when holdout behavior stops improving.

Budget Pine runtime

Control chart history

Set calc_bars_count to a deliberate value. History length, dataset capacity, sequence length, batch size, hidden width, and training frequency multiply one another.

Gate expensive work

Use .withTrainingGate(true) for heavy models. If you need history-wide adaptation, combine .trainEveryCall() with confirmed-bar, stride, and recent-window conditions.

Keep advanced blocks compact

Use short sequences, low filter/unit counts, and few attention heads. Pine Script is a chart runtime, not a GPU environment.

Reproducibility

• Seed weight initialization with nl.rng(seed).
• Keep feature ordering and hyperparameters explicit.
• Use deterministic replay seeds.
• Record the chart history window used for training.
• Export weights when a stable research snapshot matters.

Limitations

• Execution is bounded: models must remain compact enough for TradingView limits.
• History is contextual: bounded training windows can produce different state when reloaded later.
• Hyperparameter changes rebuild state: a changed script input can restart learning.
• Distribution shifts happen: validation on one regime does not guarantee another.
• Outputs are research signals: NeuraLib does not model your risk, execution, or financial objectives automatically.