Methodology
Two years. Three commodities. 7,685 predictions.
We back-tested our options-implied probability methodology against realized spot from 2024-01-02 → 2026-05-14, evaluating five synthetic strikes (95% / 98% / 100% / 102% / 105% of spot) at a 7-day horizon for silver, gold, and oil. Results are out-of-sample — we re-built the IV smile from each day's historical options chain and computed the probability fresh, then resolved it against what spot actually did.
What you want from this chart: monotonicity. A bar that says "predicted 60%" should resolve YES more often than one that says "predicted 30%". The absolute level is informative too — when predicted and realized track the diagonal, our model is well-calibrated; when they diverge, it tells you something about the regime.
7,595 of 7,685 predictions resolved (98.8%). Remainder are recent dates whose 7-day horizon has not yet landed.
Summary by commodity
| Commodity | Predictions | Resolved | YES hit rate | Mean predicted |
|---|---|---|---|---|
| silver | 2,515 | 2,485 | 59.6% | 49.6% |
| gold | 2,425 | 2,395 | 58.0% | 50.7% |
| oil | 2,745 | 2,715 | 55.1% | 49.7% |
Calibration plots
Each decile pairs the mean predicted probability against the realized hit rate for predictions that fell in that bucket. Dashed diagonal is perfect calibration.
How to read the bias
Silver and gold systematically under-predict in the middle buckets — our 50% says ~66% realized. This is the classic signature of a bull regime. Black-Scholes prices off the risk-neutral drift (r − q); realized drift over the 2024-2026 window was much higher than that, so YES outcomes happened more often than the risk-neutral measure suggested.
Oil is the tightest — it traded sideways over the same window and the model lands almost on the diagonal. That's the cleanest evidence the methodology itself is well-shaped: the bias scales with the underlying's realized drift, not with anything our process is doing wrong.
All three are monotonic across every bucket: a higher predicted decile resolves higher. That's the table-stakes claim — when you see us publish a 70% read on a Kalshi market, you should trust it as a 70% probability under our model, with absolute bias bounded by realized drift.
Methodology disclosure
- Synthetic strikes (0.95 / 0.98 / 1.00 / 1.02 / 1.05 × spot) — no historical Kalshi quote comparison because Kalshi prunes resolved-market trade history past ~6 weeks.
- 7-day horizon, resolved against next available trading-day close.
- Underlying spot from ETF proxy (SLV/GLD/USO) daily close — not the commodity futures fix.
- IV solved per-strike via Brent root-finding from historical options OHLCV-1d close prices, then linearly interpolated across the smile to the synthetic strike.
- Risk-free rate fixed at 4.5%, dividend yield 0% — close to the 2024-2026 average but the backtest is not sensitive to small shifts in either.
- Source: real-time market analysis from a licensed institutional-grade options feed, consolidated US NBBO. All historical pulls comply with the personal-use data license.