PolyEdgeFinder vs SpectralEdge

What SpectralEdge does

Signal service for Kalshi's Bitcoin 15-minute markets, from Cotes AI, Inc. Every 5 seconds a model re-reads each open market and calls YES or NO only when its calibrated confidence clears a fixed threshold and the price still leaves a minimum edge; about half of all markets get no signal. Each signal shows its side, confidence, entry price and edge, and is graded at settlement. You place trades yourself on Kalshi, and the service never connects to your account. The same feed is available as a JSON and streaming API. It starts with a free trial; the subscription price is not published.

• A fresh read of each market every 5 seconds, with confidence calibrated separately for each point in the market
• Signals carry side, confidence, entry price and estimated edge, and link straight to the Kalshi market
• Only the first signal in a market is scored, against Kalshi's settlement
• Public replays of the last ten settled markets, with no sign-up
• REST and Server-Sent Events API with graded settled history and an OpenAPI contract

• The headline is accuracy, and the company itself says accuracy alone proves nothing: its own guides explain that a 90¢ contract needs to win more than 90% of the time, and its notes for AI assistants ask that any win rate be quoted with the average entry price. It does not publish that average. Its public replay feed showed 1,220 wins and 199 losses over the trailing 30 days on Nov 5, 2026, or 86.0%, across about half of the markets in that period, in line with the 85.8% on the home page.
• The last ten markets in that feed give a first look at price. Seven drew a signal, all seven won, and they entered at 81¢ to 92.6¢, averaging about 88¢. All ten settlements matched Kalshi's public API. Seven signals are far too few to judge, but an 86% win rate at an 88¢ average entry would lose money even before Kalshi's fee, so the number to check is the average entry across the full history. Trial users can compute it from the API's entry price and outcome fields. The signal also has to be tradable: you act by hand, in a market that can reprice within a minute.
• The sample sizes vary from page to page: 15,000+ markets of out-of-sample validation, a backtest of 8,085 first signals at 88.5%, “Trained & tested on 7,654 markets” on the parent site, and more than 30,000 settled markets in a founder interview. They may describe different stages, but the site does not reconcile them. The backtest is simulated and carries the CFTC hypothetical-performance notice. “Physics-trained AI” here means decomposing Bitcoin's price into component waves, which the company says is covered by five provisional patent applications; we did not verify that.