The scorecard, updated monthly.
Every month we take our forecast from a year ago and compare it to what actually happened. Below is how far off we were, and whether we beat the simplest possible baseline — “just assume prices keep doing what they’ve been doing.” If we can’t beat that, we’re not useful. We post it either way.
First backtest run pending.
The monthly backtest cron runs on day 28 of each month. The baseline score lands as soon as we have 12 months of forecast history; the per-chain decomposition score lands once enough rate-stages have real data.
What the numbers mean
The number on the left is our “error” — how far off our forecast was on average. Lower is better. 5% would be excellent; 10% would be mediocre. The number on the right is how far off the simple baseline was, tested against the exact same months. If our number is lower, we’re adding signal.
Why show two windows
The first uses every year of history to make the forecast. The second only uses the most recent three years. If the second one is much worse, it means we’d be over-fitting to patterns that aren’t coming back. Seeing both tells you whether the forecast is durable or just lucky.
Don't trust us — verify it
These numbers come from /api/v1/performance. Pull it, graph it, compare it to whatever other forecast you use. No password required. If the numbers surprise you, email us.