Why a single accuracy number misleads
Late in a market, the likely winner is usually priced high already: three minutes before close, YES bids sat at 95¢ or above more often than in any other 5¢ band. Calling the favorite at that point can be right more than 9 times in 10 and still lose money, because the market already charges 95¢ or more for it. The accuracy that matters is accuracy measured against the price at the moment of the call.
The market is already a prediction
A YES price of 70¢ is the market saying "about a 70% chance". Across 9,312 settled BTC 15-minute markets from July 17 to October 23, 2026, every 5¢ price band won within about 2 points of its price.
| YES bid | YES won |
|---|---|
| 50–55¢ | 51.8% |
| 70–75¢ | 73.4% |
| 90–95¢ | 93.8% |
| 95–100¢ | 99.1% |
So the bar for a prediction isn't 50%. It's whatever the price already says. The full table is in how often 90¢ contracts win.
How to read a track record
- Price at the time of the call. A record should show what each signal would have cost, not just whether it won.
- Recorded before settlement. Results should be logged when the call is made and scored automatically, with nothing removed afterwards.
- Sample size. With a new market every 15 minutes, a few days can look great or terrible by chance alone. Look for hundreds of markets at least.
- Live versus backtest. A backtest is a simulation on past data. Treat live, recorded results as the stronger evidence, and expect them to differ.
- Profit after fees. If a tool shows only accuracy, apply the break-even rule yourself: win rate has to beat the average price paid. See why a 90% win rate can still lose money.
Can anyone predict 15-minute moves?
Not with certainty, and a claim of certainty is a red flag. What a model can do is estimate a probability that is better calibrated than the current price at some moments and stay quiet the rest of the time. Those moments are where an edge exists, if one exists at all.
SpectralEdge's record
SpectralEdge publishes a live, trailing 30-day record of its first signal in each market, recorded before resolution and scored automatically, alongside an out-of-sample backtest. Both are on the home page with the method and required disclosures. Each signal also carries the price and the edge at the moment it fired, so you can check it against the break-even rule. How SpectralEdge decides covers the thresholds.