Track a Kalshi Portfolio

From Forecasts to a Portfolio

In our companion blog post, we built an AI forecasting pipeline that researches Kalshi prediction markets and produces probability estimates for each one. The natural follow-up question: are the forecasts actually accurate?

Prediction markets give us a hard benchmark. If the AI's probability estimates are better than the crowd's, a portfolio that trades on them should make money — buying YES when the forecast is above the market price, and NO when it's below. If the portfolio loses money, the forecaster isn't adding value. If it makes money, that's strong evidence the AI is producing genuinely useful probability estimates.

This case study describes how we construct and track that portfolio.

Disclaimer: This is not investment advice. The portfolio described here uses simulated allocations for the purpose of benchmarking forecast accuracy. No real trades are placed.

How the Trader Works

The trader notebook takes the forecaster's output CSV and builds a simulated portfolio in four steps:

1. Fetch order books

For each forecasted market, we pull the live Kalshi order book — not just the midpoint price, but the full set of resting limit orders on both sides. This tells us what prices are actually available and how many shares we could buy at each price level.

2. Filter for edge

Not every disagreement between our forecast and the market is worth trading. We filter for markets where:

• The edge (difference between our forecast and the market price) is at least 2%
• The expected return, annualized, exceeds our threshold
• The market resolves soon enough that the return on capital is attractive — shorter-duration markets mean faster compounding

This keeps us out of marginal bets where the forecaster barely disagrees with the market.

3. Set target positions

We divide the $100,000 portfolio equally among all qualifying markets. For each position, we walk the order book, accumulating shares only at prices within our edge filters. If we're buying YES, we take asks up to our maximum price; if we're buying NO, we take asks on the other side.

4. Record what fills

Not all target positions fill completely. Thin order books, wide spreads, and prices outside our filters all reduce fill rates. We record the shares bought, average price paid, and fill percentage for each position. This is a realistic simulation — we're limited by actual available liquidity, not by wishful thinking about what we could buy.

The Portfolio: February 26, 2026

From 153 forecasted markets, the trader identified 24 positions with sufficient edge. Here's what the simulated portfolio looks like:

The portfolio is only about half deployed — a consequence of taking real order book liquidity seriously. Many markets have thin books where you simply can't get the size you want at acceptable prices. We'd rather be half-invested at good prices than fully invested at bad ones.

Here are all 24 positions, sorted by amount invested:

The portfolio spans politics (Texas Senate, Florida Governor, California Governor), geopolitics (Greenland, Khamenei, Starmer), economics (rate cuts, CPI, gas prices, marijuana rescheduling), AI (top-ranked models), and policy (DHS funding, Warsh nomination). This diversity is a feature — it means our accuracy isn't dependent on getting one domain right.

The eight fully filled positions ($4,167 each) are the backbone of the portfolio, accounting for two-thirds of deployed capital. These are markets with enough liquidity to absorb the full allocation. The remaining sixteen positions fill partially, from 76% down to essentially zero — the order books just didn't have enough shares at acceptable prices.

Some notable positions:

• Greenland NO at $4,167 — The AI thinks there's only a 10% chance the US takes control of any part of Greenland before January 2027, versus the market's 41%. That's our largest edge at +31 points.
• Aliens NO at $4,167 — Market says 23% chance the US confirms aliens exist before 2027; our forecast says 8%. Fifteen points of edge on a question where the base rate for government alien confirmations is, historically, zero.
• OpenAI top-ranked model YES at $4,167 — The AI thinks there's an 85% chance OpenAI will have a #1 ranked AI model this year, versus the market's 59%. The research cites OpenAI's track record and upcoming model releases.
• **SpaceX >12 launches YES at $60** — Our biggest edge (+39 points) but almost no fill. The order book was nearly empty, so we got only$60 deployed despite a massive forecast disagreement.

What Success Looks Like

A 30% annualized return would be a remarkable achievement — that's about 2.2% per month. We'll run the forecaster weekly, update positions, and track the portfolio's mark-to-market value over time.

This is an empirical question. We don't know yet whether the AI forecaster adds enough accuracy over the crowd to generate returns. But the structure is set up to tell us clearly: the portfolio goes up if our forecasts are better than the market, and down if they're worse. No ambiguity, no hand-waving — just a P&L that reflects forecasting accuracy.

We hope the signal is clear within a month or two. If the portfolio appreciates, that's evidence the forecaster is doing something useful. If it doesn't, we'll know we need to improve the pipeline.

Follow Along

• Forecaster blog post — How we produce the forecasts
• Full forecast data — Research, rationales, and forecasts for all 153 markets
• Kalshi Forecaster notebook — Run the forecaster yourself
• Kalshi Trader notebook — Run the trader yourself

We'll update this page as we track the portfolio over the coming weeks. We hope you enjoy following along!