Neural Forecast Lab
The Neural Forecast Lab trains a seeded committee of small neural networks on a symbol's cycle, annual, astro, and trend features, checks the result on held-out bars and walk-forward folds, then extends the fitted line forward. Pro plan required.
Where to find it
Open Dashboard → Charts → Analysis → Neural Forecast Lab. The applied overlay is the Neural Forecast indicator, which also lives in Indicators → Cycle Suite.
How it works
The lab reuses the same building blocks as Forecast Composer: dominant price cycles, the annual curve, an optional astro curve, and trend. A linear fit captures the broad shape first; a committee of small neural networks then learns the remaining structure and adds it on top. Because every network is seeded, the same settings and seed always reproduce exactly the same line.
Three segments appear on the chart:
- A dim training section where the model was fit.
- A dashed holdout section the model never saw during training — its tracking here is the first honesty check.
- A solid forward section extending the same fitted scenario beyond the newest bar.
Train and verify
Press Train network. The readout shows how the trained line behaved on the held-out bars — the share of bars it moved with price and the correlation of its day-to-day changes — plus a comparison against the plain linear fit, so you can see whether the neural committee actually added anything for this symbol.
The walk-forward table repeats the whole experiment on earlier windows: each fold retrains from scratch on older bars and grades the line on the bars that followed. Read it before trusting the forward segment — a scenario that only worked on the most recent window is a weak scenario. A practical reading rule: look for more than half the folds showing positive correlation together with a positive mean, and give extra weight to the most recent folds — a fit that only worked on older windows has already degraded. All numbers describe your own symbol, settings, and date range.
Apply to the chart
Apply to chart adds the Neural Forecast overlay with every setting pinned — including the resolved cycle periods and the seed — so saved layouts re-render the identical line. Use Update overlay after retraining and Remove overlay to clear it.
Settings reference
| Setting | Meaning |
|---|---|
| Hidden units | Size of each network; larger fits more shape and more noise |
| Committee | How many independently seeded networks are averaged |
| Epochs | Training budget per network; early stopping usually ends sooner |
| Seed | Reproducibility anchor — same seed, same line |
| Decay | Weight decay; higher values keep the correction smaller |
| Features | Trend, dominant cycles, annual curve, astro curve |
| Train window / Holdout / Project | Bars used to fit, bars held out for verification, bars projected forward |
Read projections responsibly
- A projected turn is a time to review price, not a predicted high or low.
- Prefer settings whose holdout and fold numbers stayed useful over a line that merely hugs history.
- Retrain after regime changes; news, gaps, and structural shifts break fitted scenarios immediately.
- Do not count Forecast Composer and Neural Forecast as two independent confirmations — they share the same underlying components.
Plan access
The Neural Forecast Lab and the Neural Forecast indicator require Pro.
Related guides
- Charts Workspace for the Analysis rail and saved layouts
- Cycle Suite for the cycle indicators and Forecast Composer
- Event Tester for event-based timing research
- Forecast Labs for the Wavelet Lab, Consensus Zones, and Swing Odds