We analysed 19 million Polymarket bets. Here is what does not work.

There is no shortage of advice about prediction markets. Follow the whales. Fade the crowd. Bet the longshots when the odds look wrong. Almost none of it comes with evidence attached.

We have been collecting every public trade on Polymarket into our own database — roughly 19 million bets, of which about 14 million are on markets that have already resolved, so we know the outcome. That is enough to stop guessing and start measuring. Here is what we found, including the parts that are inconvenient for us.

1. Copying successful traders is not an edge

This is the one that surprises people most, because "copy the winners" is the entire premise of several products.

We ranked traders by realised profit, split the timeline in half, and asked a simple question: does being profitable in the first half predict being profitable in the second half? The correlation was approximately zero. Traders in the top quartile of the first period went on to produce roughly average results afterwards. Individual accounts swung wildly in both directions.

We repeated the test filtering each trader down to only their best-performing category, and the persistence still did not appear. Past performance on Polymarket, measured properly, carries very little information about future performance.

2. Big bets do not predict outcomes

The intuition is seductive: if someone moves six figures on a market, surely they know something.

They usually do not — or if they do, the price has already absorbed it. When we bucketed bets by size, returns were essentially flat across the range. And when we isolated the genuinely anomalous trades — five times larger than the typical bet in that market — they performed slightly worse than normal-sized bets, not better.

A large bet is a large bet. It is not a prediction, and size alone is not a signal.

3. Longshots are where money goes to die

The clearest and most durable pattern in the entire dataset is the favourite–longshot bias, and it is a warning rather than an opportunity. Sorted by entry price, average return per bet looks like this:

Cheap contracts feel like lottery tickets and they price like lottery tickets: consistently overpriced relative to how often they actually win. The single most valuable thing this dataset taught us is defensive — not what to bet, but what to stop betting.

4. The favourite edge is real in history and thin in practice

Here is the uncomfortable part, and the reason we are publishing this instead of selling a signal service.

Backing favourites in the 55–85¢ band does show a positive expectation in the historical data — roughly two to three per cent per bet after subtracting a realistic execution cost, holding up on an out-of-sample split covering hundreds of thousands of trades. On paper, that is an edge.

We then ran it forward, live, for weeks across several independent accounts. Every variant converged towards break-even. Some periods showed a couple of points of profit; others gave it back. The gap between the historical number and the live result comes from the things a backtest cannot see: the spread you actually cross, the fills you do not get, and the fact that a live trigger selects a slightly different — and generally worse — set of markets than a historical filter does.

A pattern that survives a backtest has cleared the easy hurdle. Surviving contact with a real order book is the hard one.

5. The fast crypto markets are efficient

Polymarket runs short-horizon "will BTC be up or down" markets. We tested them hard, including building a fair-value model that priced the digital option directly from Binance spot data and comparing it against the market price in real time.

Over more than eleven thousand recorded signals the result was consistently negative. Near the close, when the outcome is nearly determined, the market prices it correctly. Earlier, when there is genuine uncertainty, there is nothing to exploit. These markets are efficient at the moment a retail participant can act on them.

6. Cross-market arbitrage is not retail-farmable

On multi-outcome events the YES prices of mutually exclusive outcomes should sum to one. When they do not, it looks like free money.

Usually it is not. Most apparent gaps fall into three categories: the outcome set is not exhaustive, so a winner outside the listed options makes every leg lose; the depth is a few dollars, so the opportunity cannot be sized; or the event resolves years out, which turns a five-cent gap into a poor annualised return. We ran an automated paper test for days at a realistic threshold and recorded almost nothing. Market makers keep the liquid events tight.

What we actually concluded

After all of it, the map is mostly closed. The one durable pattern is defensive — avoid longshots — and the one constructive pattern is thin enough that execution costs eat most of it.

That conclusion shaped what Polyradar is. We do not sell predictions, signals, or a promise that you will make money, because our own data says nobody should be making that promise. We build the instruments: the whale and large-trade radars, the market explorer over our historical dataset, a strategy builder with an honest backtester, and paper trading so you can test an idea for weeks before a single euro is at risk.

You bring the hypothesis. The tool tells you the truth about it — including when the truth is that it does not work.

Every figure above is a historical measurement of our own dataset, with the limits any such measurement has. None of it is investment advice, and none of it should be read as an indication of future results.