Agent-Assisted Options Screening: What I Learned Building an AI Options Tracker
Options flow data is noisy, fast-moving, and full of false signals. Here's how I built an agent-assisted screener that cuts through the noise — and what broke along the way.
Why Options Screening Is Hard
Stock screening is relatively straightforward. You define criteria — P/E under 20, revenue growth above 15%, debt-to-equity under 1 — and you filter the universe. The data is slow-moving and the signal-to-noise ratio, while not great, is manageable.
Options screening is a completely different problem. The data is:
**Enormous.** A single equity with weekly options might have 200+ contracts across different strikes and expirations. Multiply that by a universe of 500 stocks and you have 100,000+ data points updating every second.
**Contextual.** An unusual options trade doesn't mean anything in isolation. You have to know the stock's historical IV, the current earnings calendar, recent news flow, the overall market regime, and what other contracts are doing. Context is everything.
**Adversarial.** Smart money knows retail is watching options flow. Some unusual flow is genuine institutional positioning. Some is designed to mislead. Distinguishing between them is genuinely hard.
When I started building the AI Options Tracker, I thought the core problem was data access. It turned out the core problem was interpretation.
The Architecture I Landed On
The tracker has three main components:
**A data layer** that pulls options chain data for a screened universe of stocks, calculates implied volatility surfaces, and flags statistically unusual activity. "Unusual" here means volume significantly above the 30-day average for that strike/expiration combination, adjusted for overall market volume.
**A context layer** that enriches flagged contracts with relevant information: earnings date, recent price action, sector performance, any recent news. This is mostly structured data, not LLM work.
**An agent layer** that synthesizes the flagged contracts with context and generates a structured trade hypothesis. Not a recommendation — a hypothesis with explicit assumptions, risks, and conditions.
The agent is built around a simple loop: here is an unusual options contract, here is the relevant context, what might explain this positioning, and if the explanation is plausible, what would a directional trade look like?
What Actually Works
The most useful output isn't trade ideas. It's filters.
When I'm screening 50-100 flagged contracts on an active day, the agent is most valuable when it's telling me what to ignore. "This volume spike is probably earnings IV run-up, not directional positioning." "This put flow is consistent with hedging, not speculation — the stock is held widely in institutional portfolios." "This call buying coincides with a share buyback announcement from earlier this week, likely just delta hedging."
Eliminating the noise fast is worth more than generating one more hypothesis.
The agent is also useful for surfacing cross-asset context I would have missed. Options flow on a mid-cap retailer that seems random makes more sense when the agent notes that the sector ETF had significant put activity two sessions earlier. I wouldn't have made that connection manually in a timely way.
What Doesn't Work (Yet)
The agent struggles with timing. It can reason about whether a trade is directionally plausible, but it has no meaningful edge on when a thesis will play out. Options are time-sensitive instruments. A correct directional thesis with wrong timing is still a loss.
I've tried adding technical analysis context — support/resistance levels, recent momentum — and the agent does incorporate it, but I don't trust its technical analysis. It produces plausible-sounding statements about chart patterns that are hard to validate systematically.
The other limitation is calibration. The agent is confident. It generates hypotheses with appropriate hedging language, but the underlying model has no position-sizing intuition. It treats a high-conviction setup the same way it treats a marginal one. Building a calibrated confidence layer on top is an open problem.
Results (Paper-Traded)
I've been paper-trading the tracker's output for about eight months. The results are positive — roughly +11% average edge on target entries against the underlying's move — but paper trading is a generous environment. No slippage, no early assignment risk, no emotional decision-making.
The 3× faster decision cycle is real, though. The screening and context enrichment that used to take me 45-60 minutes on an active morning now takes 15. The agent output gives me a starting point for each position rather than starting from scratch.
What I'm Building Next
The next version will add a backtesting layer so I can validate screening criteria against historical options data, not just forward test. This requires access to historical IV data, which has a cost component — I'm evaluating data providers for this.
I also want to add a "regime" layer — a background model that adjusts the agent's behavior based on overall market conditions. What counts as unusual flow in a low-volatility trending market is different from what counts as unusual in a high-volatility choppy one. The current version doesn't account for this.
The longer-term vision is to fold the options tracker into the broader AI Investor Platform, so equity fundamentals and options positioning are visible in the same interface. A company's financial health and the options market's bet about its future are both inputs to a thesis. They should be next to each other.
The Broader Point About Agents
Building this has crystallized something about agent systems: they're most useful when they're doing things humans genuinely can't do fast enough, not things humans can do but find tedious.
Scanning 100 options contracts for contextual anomalies in under a minute is something humans can't do fast enough. Writing a first draft of a hypothesis is something humans can do but find tedious — and agents are less useful there, because the quality bar is high and the iteration time matters.
The best agent use cases I've found are always at the speed/volume boundary. Give me the signal I would have found anyway if I had more time. That's valuable. The interpretation can still be mine.