Skip to content

Cross-asset and macro market signals

Issue #36 adds a small, reproducible risk-context feature set without making the production forecast depend on external assets.

Signals

Crypto-beta context comes from completed Kraken ETHUSD hourly candles aligned to the exact BTC forecast timestamps. The derived features are ETH 1h/6h/24h returns, ETH-vs-BTC relative strength at 6h/24h, and rolling BTC/ETH return correlations over 24h and 168h.

Macro risk context comes from public FRED series:

  • VIXCLS: VIX close and five-observation percentage change;
  • DGS10: US 10-year Treasury yield and five-observation change in basis points.

Daily macro observations are conservatively considered available only at 00:00 UTC on the following calendar day. This intentionally sacrifices some freshness so backtests cannot accidentally use a same-day close before it was published. Values older than seven days are treated as stale.

Production behavior

The production run fetches ETH and the two FRED series after the BTC market-data selection. Provider errors are isolated. Any available values are added to immutable market_features, while forecast.json and the experiment manifest retain provider, freshness, missing-feature and error diagnostics.

These signals do not alter production predictions or ensemble weights merely because they are available. A missing ETH or macro provider therefore cannot stop the forecast or X publication.

Walk-forward ablation

cross_asset_ablation.py uses the recent bounded Kraken BTC/ETH history and FRED observations. At each simulated origin:

  1. BTC feature engineering sees only candles available at that origin;
  2. ETH rows newer than that origin are ignored;
  3. FRED rows are filtered by the next-day availability rule;
  4. a training example is eligible only after its target horizon has matured.

The same ridge forecaster is evaluated with the core spot features and with core + cross-asset/macro features on identical origins. Results are reported separately for 2h, 4h, 8h and 16h, include regime breakdowns, direction accuracy and deterministic paired-bootstrap uncertainty.

The scheduled research workflow retains the report for 90 days. A feature group is never promoted solely from in-sample fit; promotion is left to the feature-selection policy introduced by issue #37.