Forecast
—
Position
Live command center for all active models. Each card shows the latest published forecast, the recommended position, and how stale the underlying data is.
Scanning model artifacts…
Run the training pipeline (Phase 41 → 42) to publish your first models.
If these predictions add value to your market analysis, please consider supporting the project so it can remain accessible and free for everyone. Designing, maintaining, and hosting these models on cloud infrastructure involves ongoing costs.
Disclaimer: This dashboard is provided for informational purposes only and does not constitute financial advice.
Forecast
—
Diagnostic only
Phase 42 did not promote this run for live trading. Open the run inspector for the full audit.
Position
Loading horizon context…
Loading Model Runs…
Phase 42 · Adjudicated Runs
Investigate the published runs behind the live signal: what passed, how it performed out-of-sample, what features the champion leaned on, and where it breaks. Every panel reads the same Phase 42 snapshot.
Target → run timestamp.
Identity strip for the loaded snapshot.
—
Columns are split into three groups: Forecast quality (model error on the prediction itself; always shown), Investment economics (cost-aware Phase 41 strategy metrics — the same numbers Phase 42's alpha-vs-buy-and-hold gate uses; tradable targets only), and Robustness (multiple-testing-adjusted Sharpe + backtest-overfit probability on the cost-aware strategy return series; tradable targets only). Hover any column header for the formula and source phase.
| Investment economics · Phase 41 cost-aware strategy | α vs Buy-and-Hold · Phase 42 F3 gate | Forecast quality · OOS predictions | Robustness · cost-aware return series | Forecast error | vs Persistence baseline | |||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Model | Role | Last OOS Prediction | Annual Return | Sharpe† | Max Drawdown | Sortino† | Calmar | Annual Vol | α vs BH | Sortino edge | MDD edge | Hit Rate | Direction p¶ | PSR‡ | DSR‡ | PBO§ | RMSE | MAE | Inc. rIC vs Persistence | RMSE vs Persistence |
Hit Rate — share of OOS dates where sign(pred) matches sign(y).
Direction p — Pesaran–Timmermann one-sided p-value on directional accuracy.
RMSE / MAE — error on the raw forecast vs the realized target, no execution path.
All sourced from Phase 41 oos_metrics.parquet's strategy_* columns — cost-aware, threshold-aware, lag-aware execution path. For multi-month-horizon targets the strategy is the overlapping-sleeve portfolio using 1-month period returns; for 1-month targets the period return is y.
Annual Return = geometric CAGR of the strategy return series. Sharpe / Sortino / Calmar / MaxDD / Annual Vol all use the same series. Identical SSOT to the Phase 42 alpha-vs-BH gate.
α vs BH = strategy Annual Return − Buy-and-Hold Annual Return. Sortino edge and MDD edge are the same differences for those two metrics. Buy-and-hold stats use the same SSOT (geometric CAGR, LPM-2 Sortino) as the strategy row, so the conventions are identical.
Scope difference on Logic C targets: the gate fires per vintage (calm OR crisis), so α vs BH reflects one sub-window's evidence. The row's Annual Return aggregates over the spliced full-OOS series (calm + crisis combined — the deployed cascade's realised path). For Logic A / Logic B targets there is only one vintage and the two columns reconcile exactly.
PSR — probability the candidate's true Sharpe > 0, computed on the cost-aware strategy return series with a serial-dependence correction for h-month overlap.
DSR — PSR adjusted for selection bias across the candidate pool.
PBO — probability of backtest overfit (Bailey/López de Prado CSCV). > 0.5 fires the executive-summary overfit warning.
For non-tradable targets there is no investable "strategy return," so PSR/DSR/PBO are omitted rather than evaluated on a misleading sign(pred) × y proxy.
Marčenko–Pastur thresholding applied during Phase 30 hierarchical clustering separates the feature-correlation eigenvalues into a noise bulk (statistically indistinguishable from a random matrix) and a structural signal. Bulk eigenvalues are replaced with their mean before linkage, so the cluster representatives are chosen from a denoised correlation rather than a sample-noise-dominated one.
Phase 30 stage clustering.health.mp_denoising. A high signal-variance share with few signal eigenvalues indicates that most of the apparent feature redundancy was random sampling noise rather than true co-movement.
Loading Signal…
Phase 50 · Signal
What the model predicts next, how much to trust it, and why. Top of page is the story; technical detail is collapsed below for the curious.
Select a target to load its latest published forecast.
Identity of the run you're looking at.
How to interpret the confidence label
The label reflects model conviction, not direction. The headline value still tells you which way; this label tells you how much to lean on it.
Three competing prediction logics are scored on the same OOS window (per-row intersection). The winner drives what fires in production; the rest are shown in the comparison table below for transparency.
Phase 42 alpha-vs-buyhold gate. Pass condition: AnnRet edge ≥ floor (primary), or Sortino edge ≥ 0 AND MaxDD edge ≥ 0 (risk-reduction override).
What each member of the deployed ensemble predicts for next period using today's features. This is the look-forward, actionable signal — the equal-weighted average of these rows is the headline forecast at the top of the page. Large vs Ensemble magnitudes flag members that disagree with the consensus.
| Model | Logic / vintage | Weight | Forecast | vs Ensemble | Composite | OOS n | Direction |
|---|
Every gate-surviving model Phase 50 scored at publish time is here. The row includes the live forward prediction and the Phase 42 common-window composite evidence so you can compare the headline against alternatives.
| Logic | Model | Vintage | Current prediction | Signal label | Composite | CI | Rank IC | Dir acc | Sortino | OOS n | Passed gates |
|---|
If these predictions add value to your market analysis, please consider supporting the project. Cloud infrastructure isn't free.