The Signal-History Dataset
Every prediction our model has published, timestamped and then resolved against the actual close — across 16 markets. It is forward out-of-sample by construction: each row was called before the outcome was known and recorded as it happened, so there is no hindsight, no survivorship cherry-picking, and no backtest-overfitting to take on trust. Download it as CSV and test it against your own ideas.
Point-in-time
Each prediction carries the date it was published. Join your own forward returns to it without leaking the future — the property professional quants check for first.
Live-resolved, honestly labelled
Only resolved rows (won / lost / no_data) are included — factual track record, not forward calls. Walk-forward backfilled rows are tagged source=backfill so you can filter them out.
Yours to analyse
Plain CSV. Backtest against it, combine it with your own factors, or score it in pandas. The signal is a continuous variable (predicted_pct), not a buy/sell instruction.
Download
Checking your access…
Schema
| Column | Meaning |
|---|---|
| date | UTC date the prediction was published (point-in-time). |
| ticker | Instrument symbol. |
| market | Market code (US, TW, AU, JP, KR, SG, HK, NZ, …). |
| horizon | Forecast horizon: 1d / 3d / 7d / 1mo / 3mo / 6mo. |
| predicted_pct | The model's predicted return — the signal, as a continuous research variable. |
| predicted_price | Predicted price implied by predicted_pct at publish time. |
| confidence | Model confidence bucket: low / medium / high. |
| cv_dir_acc | Walk-forward cross-validated directional accuracy of the model vintage (0–1). |
| actual_pct | Realised return over the horizon (filled when the row resolved). |
| actual_price | Realised close at resolution. |
| status | won / lost / no_data — the resolved outcome. Only resolved rows are in this file. |
| source | live = published in real time; backfill = walk-forward simulation (no look-ahead). Labelled honestly so you can filter. |
| verified_at | Timestamp the row was resolved against the actual close. |
Resolved rows only. This dataset contains predictions whose horizon has elapsed and been scored. It is a historical record, not a set of current recommendations, and is not investment advice.
Coverage is honest. The archive is young — it reflects predictions since the model went live, plus clearly-labelled walk-forward backfill. Treat statistics from a few months of data as preliminary; significance grows with the archive.
Personal / research use.Please don't redistribute the raw file or resell it as your own dataset.