What I Learned Trying to Predict Polymarket Bets

What I Learned Trying to Predict Polymarket Bets

Phase 1 of the Polymarket work: build the dataset, enrich it properly, and discover that collection was harder than modeling.

I started with the most intuitive idea: predict whether a binary Polymarket market was mispriced before resolution. The real surprise was not the modeling challenge. It was how much engineering was needed before the dataset even existed.

This phase came before the 15-minute bot and before the HFT recorder. It set the rule that shaped everything afterwards: own the data path first, then decide if the research deserves more time.

If the dataset does not exist, data collection becomes the real project.

The thesis was broader than simple price prediction. I wanted to combine:

• trade flow and price buckets from the Data API,
• CLOB structure such as spread, midpoint, and depth,
• market metadata such as duration, category, and volume,
• holder-level signals such as whale concentration, trader skill, and side lean.

That last part matters. The scraper was not only gathering trades. It also had a full optional layer for top holders, trader metrics, and market lean weighted by whale score.

This first phase already contained signals that later justified the whole portfolio story.

• execution_mode supported TRAINING and LIVE.
• min_volume_usd and initial_offset controlled scope from the config layer instead of being hardcoded.
• top_holders_limit went up to the API cap of 20 holders.
• Trader metrics were cached for 5 days so the pipeline did not waste calls recomputing wallet histories.
• Whale influence was summarized into whale_score, then reused to compute lean_yes_pct and lean_no_pct.

That means the project was not just "download trades and train a model." It was already trying to express market structure, participant quality, and operational robustness.

The blocker was not feature engineering. It was the cost of collecting enough history to trust the research.

A few thousand events could already mean hours of work. Tens of thousands of clean examples meant many long runs, more storage, and much more operational babysitting than I wanted for a first research front.

The pause was not a verdict on prediction markets as a topic. It was a scope decision.

• Long-horizon markets resolve slowly, so the learning loop is slow.
• Coverage had to be broad enough to avoid training on a thin or biased slice.
• The richer I made the dataset, the more collection became the bottleneck.

Technically, the phase was successful. Strategically, it was too expensive relative to the feedback it was giving me.

This phase still paid off because it clarified what the next phases should look like.

• It pushed me toward 15-Minute Trading on Polymarket, where the loop was much faster.
• It made the case for HFT on Polymarket: Model, Rust, and the 98% Lie, where recording my own data became practical.
• It reinforced the realism mindset that later became How a Real Backtest Works.

It also left behind bundle tooling and a clean module layout, which mattered later when the project had to become portable and server-friendly.

Prediction on Polymarket was technically viable, but the first serious obstacle was not the model. It was building a trustworthy dataset rich enough to deserve one.

That lesson shaped the rest of the journey: shorter horizons, tighter loops, and a much clearer respect for the difference between a good idea and a sustainable data problem.