How does SpectralEdge work? An interview with its founders

SpectralEdge by Cotes applies signal-processing methods to short-duration prediction markets. We spoke with co-founders Steven Lacy, CEO, and Michael Gauthier, CTO, about the system's physics-based inputs, its live record, and how traders should interpret its signals.

This interview was prepared with PM Wisdom. For an independent product overview, links, and risk notes, see the SpectralEdge by Cotes listing on PM Wisdom.

What pushed you to build SpectralEdge?

Steven Lacy: I began applying spectral decomposition to markets in early 2025. The method treats price like a physical signal: break it into component waves, then measure how those waves line up. I started with equities, moved into crypto, and found that the approach transferred across assets and market structures.

Prediction markets were taking off around the same time. Mike introduced me to Kalshi and Polymarket, and BTC 15-minute markets gave us an unusually fast environment for testing the method. There are 96 contests a day, each tied to a liquid underlying asset. That let us record thousands of resolved markets, train the model, and test live signals on a schedule that would take years in many traditional markets.

After roughly 15,000 contests of live signaling, we believe the method has demonstrated enough consistency to support a real product.

The risk side is personal for me. I have five children, and prediction markets made it obvious how quickly someone could lose a meaningful amount of money. Many tools give a buy or sell signal at the beginning and then disappear. SpectralEdge keeps reading the market throughout the contest. If an initial up signal turns down, that change appears on the tape so the trader can decide whether to reduce risk. Users can also set their own confidence and edge thresholds.

Why start with BTC 15-minute prediction markets?

They give us a large amount of evidence quickly. A new market opens every 15 minutes, so we have been able to validate the system against more than 30,000 settled markets instead of waiting months for a smaller sample to resolve.

Every call is graded within minutes, which makes the live record easy to inspect. The market structure also fits the engine. Kalshi's BTC 15-minute series, KXBTC15M, asks whether BTC's 60-second average on the CF Benchmarks BRTI will finish above the strike. That is a short-timescale question tied to a live underlying price, which is exactly the kind of signal our model reads.

What does "physics-trained AI" mean in practice?

It means the model starts with signal physics instead of conventional chart indicators. The process has four parts:

1. Spectral decomposition breaks the live BTC price into component waves operating at different timescales.
2. Resonance measures how closely those waves align. Waves moving together may sustain a move, while waves cancelling one another may not.
3. Cross-feed analysis reads BTC and the prediction contract together, giving the model information about both the underlying asset and the market's current view.
4. Calibrated confidence converts those readings into a probability. A well-calibrated 90% estimate should resolve correctly about 90% of the time across a sufficiently large sample.

The mathematics can be applied to other time series and timescales, which is why we see SpectralEdge as an engine rather than a single-market model. We have filed five provisional patent applications covering the method.

What makes a signal strong enough to publish?

It must pass two thresholds that we set in advance and apply consistently.

First, confidence must clear the selected threshold. We use separate isotonic calibration curves for up and down outcomes at every five-second point in the market. A model reading at minute three is evaluated against its own calibration, as is a reading at minute twelve.

Second, the gap between our confidence and the contract price must be large enough to constitute edge. Being 97% confident in a side priced at 97 cents does not leave useful edge, so we would not publish that as an actionable signal.

Signals are recorded when they fire, before resolution. We do not select winners after the fact, and some markets receive no signal. We also stop issuing signals during the final two minutes. Those late readings may be highly accurate, but prices are usually too close to resolution value to leave worthwhile edge.

How do price, confidence, and edge work together?

Price is what the market charges for a side and therefore its implied probability. A YES contract at 70 cents implies a market probability of roughly 70%. Confidence is our calibrated probability for the same outcome. Edge is the difference between the two.

If SpectralEdge assigns an 85% probability to YES while the contract trades at 70 cents, the estimated edge is about 15 cents per contract. If the contract already costs 88 cents, our estimate is below the market price and there is no positive edge under that reading.

This is why win rate cannot be read in isolation. Buying contracts at 90 cents and winning 89% of the time still loses money. Every SpectralEdge signal includes its direction, confidence, entry price, and estimated edge so traders can judge whether the price available to them still supports the trade.

Why does the model re-read each market every five seconds?

Both parts of the decision change during the contest. BTC moves, which changes the wave structure and the model's confidence. The contract price also moves, which changes the available edge.

In a 15-minute market, a setup can appear and become fully priced within a minute. Reading the market every five seconds lets SpectralEdge record the first point at which both confidence and edge clear their thresholds. It also gives traders a changing view of the setup instead of leaving them with one stale call.

Each reading is evaluated against the historical sample used to validate that stage of the model.

How should users interpret the live accuracy record and backtest?

We record every signal. Subscribers and free-trial users can inspect the live history in the app, and developers can access it through the API.

At the time of this interview, the app showed trailing 30-day first-signal accuracy of 87%. That figure had increased after we tightened the calibration in September.

The out-of-sample backtest covers more than 15,000 consecutive markets using the same first-signal policy and time slots used live. It produced first-signal accuracy of 88.5% with coverage slightly above 50%, meaning the system issued a signal on roughly half of the markets and sat out the rest.

Accuracy still needs to be considered alongside entry price. A high hit rate on expensive contracts can lose money, which is why the product presents edge as well as direction.

What can developers build with the JSON and SSE APIs?

The API provides the same core signal data shown in the app: side, confidence, entry price, edge, and graded outcome.

• GET /v1/contests/active returns the live market and its signals.
• GET /v1/contests returns settled history with graded outcomes, allowing developers to test their own rules against SpectralEdge signals.
• GET /v1/stream/contests provides a server-sent events stream for contests and signals.

Developers can use this data for alerts, dashboards, research, or trading systems with their own sizing and execution rules. SpectralEdge also provides an OpenAPI contract, an llms.txt file, and setup prompts for Claude Code and Cursor. The SpectralEdge API documentation covers the current endpoints.

How should traders use SpectralEdge signals responsibly?

A signal is a calibrated probability, not an instruction. Some signals will lose, and losing streaks are normal in any probabilistic system.

Traders should check the price currently available to them. If the market has moved beyond the signal's entry price, the original edge may no longer exist. Position size should also be small enough that a losing streak remains manageable, and there is no obligation to act on every signal.

Users place trades in their own Kalshi accounts. SpectralEdge does not connect to those accounts, and Cotes does not custody or move customer funds.

The signals are research output shared equally with subscribers. They are not financial advice, and event-contract trading carries a real risk of loss.

What are you building next?

We are expanding into more liquid prediction markets. The engine is not limited to Bitcoin or 15-minute intervals, so the same signal stack can support other markets and timescales.

Our goal is to build one engine that can serve human traders, trading agents, desks, and the platforms they use.