Model performance (Perp Prediction)

This page summarizes recent prediction performance (last 7 days of 10‑minute snapshots). We focus on short-horizon directional accuracy and probability calibration.

Quick answers (AEO/GEO)

Summary stats (last 7 days)

Evaluated predictions804
Accuracy49.3%
Precision (UP)51.8%
Recall (UP)52.0%
Brier score (lower is better)0.2519
Log loss (lower is better)0.6970
Avg |realized return| (10m)+0.22%
Avg hourly sigma (log std)0.0083

Calibration (P(UP) buckets)

If probabilities are calibrated, then when the model says “60–65%”, UP should happen about 60–65% of the time (in that bucket).

P(UP) bucketNEmpirical UP rate
0–55% 775 53.3%
55–65% 29 34.5%
65–75% 0
75–100% 0

Recent narrative (mechanism + news context)

Large BTC moves are often accompanied by rising realized volatility, outsized |10‑minute returns|, and frequent direction flips—conditions where simple momentum/mean‑reversion signals can become unstable.

To interpret why a specific spike occurred, cross‑reference spike timestamps in the archive with reputable, time‑stamped sources such as:

Practical workflow: find the largest |10‑minute returns| windows → search those timestamps in the sources above → annotate the likely catalyst(s), and whether the move looks like a news shock vs a positioning/liquidation cascade.

Not financial advice. Educational analytics only.