Bridging the NLP Gap: How I Built a Thesis-to-Trade Translator for Prediction Markets

Prediction markets like Kalshi are genuinely interesting instruments. Unlike traditional financial markets, they price discrete binary outcomes — will the Fed cut rates in Q2? Will CPI come in above 3%? The contracts are clean, the payoffs are defined, and the information embedded in prices is surprisingly rich.

But there's a UX problem that I kept bumping into as someone who actually trades on these platforms: the gap between forming a thesis and executing a structured trade is large, and it's filled with friction.

I'd read a Fed statement, form a clear opinion — "the market is pricing in 60% odds of a May cut and that's too high given the language around persistent services inflation" — and then stare at Kalshi trying to figure out exactly how to express that as a trade. Which contract? What direction? How much? When do I exit?

The Problem Is Translation, Not Intelligence

This isn't really an intelligence problem. Most active traders on prediction markets have opinions. They have reasons. What's hard is the translation step — going from a natural language thesis to a structured, auditable trade spec.

That translation involves:

• Market identification: Finding the specific contract that best expresses the thesis
• Directional clarity: Is a YES or NO position the right vehicle?
• Sizing logic: Given conviction level and account size, what's an appropriate position?
• Entry conditions: Is now the right time, or should you wait for a trigger?
• Exit logic: What outcome invalidates the thesis early? What's the target?

Most traders do this in their head, under time pressure, and make at least one of those decisions poorly.

What I Built

PredictScript is a single-input tool that takes a plain English thesis and returns a structured trade spec covering all five of those dimensions. Under the hood, it uses an LLM to parse intent from natural language — extracting the underlying economic view, identifying the relevant Kalshi market category, inferring directional stance, and generating entry/exit logic that's consistent with the stated thesis.

The output isn't a recommendation. It's a translation. The goal is to hand you back something you can actually review and audit before executing, rather than a gut-feel decision made on the fly.

The NLP Challenge

The interesting engineering problem is ambiguity resolution. A thesis like "fade the cut pricing" assumes domain knowledge (Fed funds futures, FOMC language, rate expectations). A thesis like "I think the Lakers miss the playoffs" is much more direct. The model has to handle both ends of that spectrum without hallucinating markets that don't exist or producing sizing logic that ignores stated conviction levels.

I'm using structured output with a defined schema so the response is always parseable — market category, contract name pattern, direction, size tier (small/medium/large based on stated conviction), and a plain-English rationale for each field. That schema enforcement has been more valuable than any amount of prompt engineering for making outputs actually usable.

Where It Is Now

It's live and working at https://getpredictscript.com. It's early. The market lookup isn't connected to live Kalshi data yet, so the output is a spec you'd then go execute manually rather than a one-click trade. That's next.

If you trade on prediction markets and want to try it, I'd genuinely value feedback on what the output gets wrong or what's missing from the trade spec structure.