Backtesting
Point-in-time history and how to test without look-ahead bias.
Use live history for honest tests
Add live_only=true to GET /v1/metrics/{id}/timeseries or GET /v1/export/history. You then receive only points that were computed live at the time, from the inputs available then, and not edited afterwards. That series is point-in-time safe. It starts when the metric went live, so it is shorter than the full history.
What the provenance labels mean for a backtest
| Provenance | Safe for a backtest? | Why |
|---|---|---|
live | Yes | Computed at the time with the inputs available then. |
backfill_recomputed | With care | History recomputed with today's definition: it can embed hindsight in the formula, and inputs that are later revised (for example macro series) are the revised ones. |
backfill_approximate | With care | Reconstructed from coarser inputs; a declared approximation. |
backfill_model_retro | No | AI retrodiction under a blind protocol; hindsight leakage cannot be excluded. |
forecast | No | A projection, not an observation. |
Practical rules
- Lag your signal by one bar: a value stamped
tis available to you only aftert. - Mind definition versions:
GET /v1/metrics/{id}/definitionlists when the definition changed; do not treat a series as homogeneous across a version change. - Skip STALE periods: a metric with missing inputs publishes nothing rather than a filled value, so gaps are real gaps (Data quality).
- Include costs and slippage; metric values are indicators, not executable prices.
curl -H "X-API-Key: YOUR_API_KEY" \
"https://api.madjik.io/v1/metrics/M10004/timeseries?live_only=true&interval=1h&start=2026-09-01T00:00:00Z"Not investment advice. Past values of an indicator do not indicate future results.