I developed an application for automating and managing trades on polymarket and kalshi. About 6 months ago I started looking into whether the favorite-longshot bias from horse racing and sports betting shows up in prediction markets. It does. I pulled all 59,000 resolved binary markets off their API and ran a calibration study. Markets priced in the 40-50% range resolve Yes only about 22% of the time. Sports and games are the strongest categories.

Why I don't think it's noise

I ran Benjamini-Hochberg FDR correction at q=0.05 across 537 calibration cells. 78 survived. You'd expect about 27 by chance, so that's roughly 2.9x the false discovery rate.

Before writing any strategy code I set hard kill gates. If I couldn't find at least 8pp of miscalibration in a tradeable category, I'd stop. It passed. I also caught a Simpson's paradox artifact where the bias looked like it was growing over time but it was really just Sports becoming a bigger share of the market mix.

The part I think matters most: I tried expanding to Kalshi using 7.68 million markets and the kill gate failed. The apparent signal was a bucket-assignment artifact at the 50-cent line. I killed that track. The gates aren't decoration.

Backtest (in-sample, take it for what it's worth)

About 4,850 signals, ~150 trades after filtering. 64.6% win rate, Sharpe 1.21. Simulation from $3K to ~$8K, max drawdown 25.1%. Average hold around 20 days.

Where I'm at now

Paper trading went live this week. 12 positions open, $4K of $10K deployed. First resolutions in a week or two. I'll post an update with out-of-sample numbers when I have them, pass or fail.

I'm not sharing the specific cell map, classification system, or pipeline logic. That's the IP. But if you pulled the data yourself you'd confirm the FLB exists. Knowing it exists and knowing exactly where to trade it are different things.

What I'm curious about

• Does the methodology hold up or am I missing something obvious?
• Anyone ever commercialized capacity-constrained trading IP? The ceiling here is ~$50-100K before you move the markets you're trading. I'm deploying my own capital but I've been wondering if the methodology has standalone value as education or research.
• If you trade prediction markets, does the FLB claim match your experience?

Happy to talk process and stats. Not sharing the cell map or pipeline specifics.

Hey all. Long-time lurker, first real post. I want to share a project I have been working on and get some honest feedback — both on the methodology and on whether the IP has commercial legs.

The short version: I built a systematic trading system that exploits the favorite-longshot bias on Polymarket (CFTC-regulated prediction market). The core finding is that binary markets in the 30-60% price range are overpriced by 12-24 percentage points, and this holds up after Benjamini-Hochberg FDR correction across 59K resolved markets.

Background

Polymarket binary contracts pay $1 if an event happens, $0 if it doesn't. A contract at $0.45 implies 45% probability. If I can show the true resolution rate for that class of markets is much lower than 45%, there is a structural edge.

I collected all resolved binary markets from Polymarket's API — about 59,000 markets total. Ran a calibration study: for markets priced at X% at various time horizons before resolution, what fraction actually resolved Yes?

The favorite-longshot bias showed up clearly. Markets in the 40-50% range resolve Yes only about 22% of the time. Sports and games categories are the strongest. The bias is driven by retail traders overpaying for exciting "Yes" on longshot outcomes — the same psychological pattern that has been documented in horse racing and sports betting for decades.

Why I think this is not just data mining

This is where I expect the most pushback, so let me get ahead of it:

1. Statistical correction. I used Benjamini-Hochberg FDR correction at q=0.05 across 537 calibration cells (category x horizon x price bucket). 78 cells survived. If this were noise, you would expect roughly 27 cells to survive — getting 78 is a 2.9x multiple over the false discovery rate.

2. Pre-registered kill gates. Before writing any strategy code, I set explicit pass/fail criteria. The Phase 0 kill gate required >8pp miscalibration in at least one tradeable category. If it had failed, I would have stopped the project entirely and published the calibration study as a portfolio piece. It passed with STRONG_PASS.

3. Simpson's paradox testing. The apparent intensification of bias over time (13pp at 7 days, 24pp at 30 days) turned out to be a composition artifact — Sports grew from 7% to 26% of the market mix over the dataset period, and Sports has the strongest signal. Within categories, the bias is stable across time. I caught this with volume and category controls.

4. A kill gate that actually fired. I expanded the analysis to Kalshi (another CFTC-regulated prediction exchange) using an independent dataset of 7.68M markets. The kill gate failed — only 2 of 10 required BH cells survived, and a boundary sensitivity check revealed the apparent signal was a bucket-assignment artifact at the 50-cent line. I paused the Kalshi track based on this result. I am mentioning this specifically because it demonstrates the gates are not decoration — they fire when the signal is not there.

Backtest results (in-sample, all the usual caveats apply)

• 4,851 signals generated, ~150 trades executed through a multi-gate filtering pipeline
• 64.6% win rate, 23% ROI, Sharpe 1.21
• Post-capacity-expansion simulation: $3K starting capital to ~$8K, CAGR 63.7%, Sharpe 1.07, max drawdown 25.1%
• Average hold period: ~20 days

I am not going to pretend these are out-of-sample numbers. They are not. That is what the forward validation phase is for.

Where things stand right now

Forward validation (paper trading with live market data) went live this week. 12 open positions, about $4K of $10K budget deployed. First resolutions expected within a week or two. The system runs on 15-minute cycles with 227 automated tests and a full CI pipeline.

I do not have out-of-sample results yet. I will share an update on how forward validation went — whether it passed or failed.

What I am deliberately not sharing

I am not publishing the exact cell map (which category/horizon/bucket combinations are tradeable), the structural classification system I built for market taxonomy, or the signal pipeline gating logic. These are the core IP.

I am sharing enough of the methodology for you to evaluate whether it is rigorous, but not enough to replicate the strategy without doing the work yourself. If you ran the same calibration study on the public Gamma API data, you would confirm the FLB exists — but knowing it exists and knowing which specific cells to trade are very different things.

The commercialization question

This is the part I genuinely want community input on.

The capacity ceiling for this strategy is roughly $50-100K deployed capital before you start moving markets. That is a fundamental constraint — it means selling execution (fund, copy-trading) actively degrades the edge. But selling intelligence (methodology, data, education) does not.

The paths I am considering:

• Education: A course teaching calibration methodology and structural bias analysis for prediction markets. The techniques generalize to any prediction market, not just Polymarket.
• Research/data licensing: The 59K-market dataset with calibration results, licensed to platforms or research teams.
• Signals-as-a-service: Heavily capped (5-10 seats max) and only after 100+ forward-validated trades with confirmed edge. This is the most obvious path but also the one that erodes the moat fastest.

I have a slide deck and a detailed proposal document ready if anyone wants to discuss specifics — happy to share in a discussion with anyone who has relevant experience.

My questions for this community

1. Does the methodology sound rigorous, or am I fooling myself? What holes do you see? I have been deep in this for months and could be missing something obvious.
2. Has anyone here commercialized quantitative trading IP? What worked and what did not? I am especially interested in hearing from people who navigated the "edge is real but capacity-constrained" problem.
3. If you were shopping a slide deck for this kind of project, who would you approach? Prediction market platforms? Quant funds doing alt-data? Fintech accelerators? Educational platforms?
4. Any prediction market traders here who can gut-check the FLB claim from their own experience? Curious if this matches what you have seen in practice.

Happy to answer methodology questions. I will not share the specific cell map or signal pipeline details, but anything about the process, statistical approach, or commercialization thinking is fair game.

Introduction

Hey all. Long-time lurker, first real post. I want to share a project I have been working on and get some honest feedback — both on the methodology and on whether the IP has commercial legs.

The short version: I built a systematic trading system that exploits the favorite-longshot bias on Polymarket (CFTC-regulated prediction market). The core finding is that binary markets in the 30-60% price range are overpriced by 12-24 percentage points, and this holds up after Benjamini-Hochberg FDR correction across 59K resolved markets.

Background

Polymarket binary contracts pay $1 if an event happens, $0 if it doesn't. A contract at $0.45 implies 45% probability. If I can show the true resolution rate for that class of markets is much lower than 45%, there is a structural edge.

I collected all resolved binary markets from Polymarket's API — about 59,000 markets total. Ran a calibration study: for markets priced at X% at various time horizons before resolution, what fraction actually resolved Yes?

The favorite-longshot bias showed up clearly. Markets in the 40-50% range resolve Yes only about 22% of the time. Sports and games categories are the strongest. The bias is driven by retail traders overpaying for exciting "Yes" on longshot outcomes — the same psychological pattern that has been documented in horse racing and sports betting for decades.

Why I think this is not just data mining

This is where I expect the most pushback, so let me get ahead of it:

1. Statistical correction. I used Benjamini-Hochberg FDR correction at q=0.05 across 537 calibration cells (category x horizon x price bucket). 78 cells survived. If this were noise, you would expect roughly 27 cells to survive — getting 78 is a 2.9x multiple over the false discovery rate.

2. Pre-registered kill gates. Before writing any strategy code, I set explicit pass/fail criteria. The Phase 0 kill gate required >8pp miscalibration in at least one tradeable category. If it had failed, I would have stopped the project entirely and published the calibration study as a portfolio piece. It passed with STRONG_PASS.

3. Simpson's paradox testing. The apparent intensification of bias over time (13pp at 7 days, 24pp at 30 days) turned out to be a composition artifact — Sports grew from 7% to 26% of the market mix over the dataset period, and Sports has the strongest signal. Within categories, the bias is stable across time. I caught this with volume and category controls.

4. A kill gate that actually fired. I expanded the analysis to Kalshi (another CFTC-regulated prediction exchange) using an independent dataset of 7.68M markets. The kill gate failed — only 2 of 10 required BH cells survived, and a boundary sensitivity check revealed the apparent signal was a bucket-assignment artifact at the 50-cent line. I paused the Kalshi track based on this result. I am mentioning this specifically because it demonstrates the gates are not decoration — they fire when the signal is not there.

Backtest results (in-sample, all the usual caveats apply)

• 4,851 signals generated, ~150 trades executed through a multi-gate filtering pipeline
• 64.6% win rate, 23% ROI, Sharpe 1.21
• Post-capacity-expansion simulation: $3K starting capital to ~$8K, CAGR 63.7%, Sharpe 1.07, max drawdown 25.1%
• Average hold period: ~20 days

I am not going to pretend these are out-of-sample numbers. They are not. That is what the forward validation phase is for.

Where things stand right now

Forward validation (paper trading with live market data) went live this week. 12 open positions, about $4K of $10K budget deployed. First resolutions expected within a week or two. The system runs on 15-minute cycles with 227 automated tests and a full CI pipeline.

I do not have out-of-sample results yet. I will share an update on how forward validation went — whether it passed or failed.

What I am deliberately not sharing

I am not publishing the exact cell map (which category/horizon/bucket combinations are tradeable), the structural classification system I built for market taxonomy, or the signal pipeline gating logic. These are the core IP.

I am sharing enough of the methodology for you to evaluate whether it is rigorous, but not enough to replicate the strategy without doing the work yourself. If you ran the same calibration study on the public Gamma API data, you would confirm the FLB exists — but knowing it exists and knowing which specific cells to trade are very different things.

The commercialization question

This is the part I genuinely want community input on.

The capacity ceiling for this strategy is roughly $50-100K deployed capital before you start moving markets. That is a fundamental constraint — it means selling execution (fund, copy-trading) actively degrades the edge. But selling intelligence (methodology, data, education) does not.

The paths I am considering:

• Education: A course teaching calibration methodology and structural bias analysis for prediction markets. The techniques generalize to any prediction market, not just Polymarket.
• Research/data licensing: The 59K-market dataset with calibration results, licensed to platforms or research teams.
• Signals-as-a-service: Heavily capped (5-10 seats max) and only after 100+ forward-validated trades with confirmed edge. This is the most obvious path but also the one that erodes the moat fastest.

I have a slide deck and a detailed proposal document ready if anyone wants to discuss specifics — happy to share in a discussion with anyone who has relevant experience.

My questions for this community

1. Does the methodology sound rigorous, or am I fooling myself? What holes do you see? I have been deep in this for months and could be missing something obvious.
2. Has anyone here commercialized quantitative trading IP? What worked and what did not? I am especially interested in hearing from people who navigated the "edge is real but capacity-constrained" problem.
3. If you were shopping a slide deck for this kind of project, who would you approach? Prediction market platforms? Quant funds doing alt-data? Fintech accelerators? Educational platforms?
4. Any prediction market traders here who can gut-check the FLB claim from their own experience? Curious if this matches what you have seen in practice.

Happy to answer methodology questions. I will not share the specific cell map or signal pipeline details, but anything about the process, statistical approach, or commercialization thinking is fair game.

Introduction

Hey all. Long-time lurker, first real post. I want to share a project I have been working on and get some honest feedback — both on the methodology and on whether the IP has commercial legs.

The short version: I built a systematic trading system that exploits the favorite-longshot bias on Polymarket (CFTC-regulated prediction market). The core finding is that binary markets in the 30-60% price range are overpriced by 12-24 percentage points, and this holds up after Benjamini-Hochberg FDR correction across 59K resolved markets.

Background

Polymarket binary contracts pay $1 if an event happens, $0 if it doesn't. A contract at $0.45 implies 45% probability. If I can show the true resolution rate for that class of markets is much lower than 45%, there is a structural edge.

I collected all resolved binary markets from Polymarket's API — about 59,000 markets total. Ran a calibration study: for markets priced at X% at various time horizons before resolution, what fraction actually resolved Yes?

The favorite-longshot bias showed up clearly. Markets in the 40-50% range resolve Yes only about 22% of the time. Sports and games categories are the strongest. The bias is driven by retail traders overpaying for exciting "Yes" on longshot outcomes — the same psychological pattern that has been documented in horse racing and sports betting for decades.

Why I think this is not just data mining

This is where I expect the most pushback, so let me get ahead of it:

1. Statistical correction. I used Benjamini-Hochberg FDR correction at q=0.05 across 537 calibration cells (category x horizon x price bucket). 78 cells survived. If this were noise, you would expect roughly 27 cells to survive — getting 78 is a 2.9x multiple over the false discovery rate.

2. Pre-registered kill gates. Before writing any strategy code, I set explicit pass/fail criteria. The Phase 0 kill gate required >8pp miscalibration in at least one tradeable category. If it had failed, I would have stopped the project entirely and published the calibration study as a portfolio piece. It passed with STRONG_PASS.

3. Simpson's paradox testing. The apparent intensification of bias over time (13pp at 7 days, 24pp at 30 days) turned out to be a composition artifact — Sports grew from 7% to 26% of the market mix over the dataset period, and Sports has the strongest signal. Within categories, the bias is stable across time. I caught this with volume and category controls.

4. A kill gate that actually fired. I expanded the analysis to Kalshi (another CFTC-regulated prediction exchange) using an independent dataset of 7.68M markets. The kill gate failed — only 2 of 10 required BH cells survived, and a boundary sensitivity check revealed the apparent signal was a bucket-assignment artifact at the 50-cent line. I paused the Kalshi track based on this result. I am mentioning this specifically because it demonstrates the gates are not decoration — they fire when the signal is not there.

Backtest results (in-sample, all the usual caveats apply)

• 4,851 signals generated, ~150 trades executed through a multi-gate filtering pipeline
• 64.6% win rate, 23% ROI, Sharpe 1.21
• Post-capacity-expansion simulation: $3K starting capital to ~$8K, CAGR 63.7%, Sharpe 1.07, max drawdown 25.1%
• Average hold period: ~20 days

I am not going to pretend these are out-of-sample numbers. They are not. That is what the forward validation phase is for.

Where things stand right now

Forward validation (paper trading with live market data) went live this week. 12 open positions, about $4K of $10K budget deployed. First resolutions expected within a week or two. The system runs on 15-minute cycles with 227 automated tests and a full CI pipeline.

I do not have out-of-sample results yet. I will share an update on how forward validation went — whether it passed or failed.

What I am deliberately not sharing

I am not publishing the exact cell map (which category/horizon/bucket combinations are tradeable), the structural classification system I built for market taxonomy, or the signal pipeline gating logic. These are the core IP.

I am sharing enough of the methodology for you to evaluate whether it is rigorous, but not enough to replicate the strategy without doing the work yourself. If you ran the same calibration study on the public Gamma API data, you would confirm the FLB exists — but knowing it exists and knowing which specific cells to trade are very different things.

The commercialization question

This is the part I genuinely want community input on.

The capacity ceiling for this strategy is roughly $50-100K deployed capital before you start moving markets. That is a fundamental constraint — it means selling execution (fund, copy-trading) actively degrades the edge. But selling intelligence (methodology, data, education) does not.

The paths I am considering:

• Education: A course teaching calibration methodology and structural bias analysis for prediction markets. The techniques generalize to any prediction market, not just Polymarket.
• Research/data licensing: The 59K-market dataset with calibration results, licensed to platforms or research teams.
• Signals-as-a-service: Heavily capped (5-10 seats max) and only after 100+ forward-validated trades with confirmed edge. This is the most obvious path but also the one that erodes the moat fastest.

I have a slide deck and a detailed proposal document ready if anyone wants to discuss specifics — happy to share in DMs with anyone who has relevant experience.

My questions for this community

1. Does the methodology sound rigorous, or am I fooling myself? What holes do you see? I have been deep in this for months and could be missing something obvious.
2. Has anyone here commercialized quantitative trading IP? What worked and what did not? I am especially interested in hearing from people who navigated the "edge is real but capacity-constrained" problem.
3. If you were shopping a slide deck for this kind of project, who would you approach? Prediction market platforms? Quant funds doing alt-data? Fintech accelerators? Educational platforms?
4. Any prediction market traders here who can gut-check the FLB claim from their own experience? Curious if this matches what you have seen in practice.

Happy to answer methodology questions. I will not share the specific cell map or signal pipeline details, but anything about the process, statistical approach, or commercialization thinking is fair game.