Research/Nov 17, 2026

Buying the late favourite in 5-minute crypto markets: no edge found

Over 1,176 windows, buying the favourite 252 seconds into a 5-minute market with a spot filter gave 14 trades, 1 reversal and -5.6%. An earlier "0 reversals" came from looking ahead.

What we asked

Polymarket runs 5-minute "Up or Down" markets on BTC, ETH, SOL and XRP. A market resolves on a 60-second time-weighted average price (TWAP) from the Chainlink oracle at the end of the window, compared with the start. The idea we tested: late in the window the crowd is usually right, so buying the favourite at 0.90 or more costs little and the favourite should finish at $1. The risk is a reversal, where the favourite loses in the last seconds. One reversal in 50 to 75 trades can erase the gain, so the test needs many windows.

An ask is the lowest price a seller will accept. Buying at the ask is a taker trade.

Data

We recorded, for each window, the spot price from Coinbase every 5 seconds, the order book for the Up outcome, taker trades and the final outcome. The earlier test covered 64 windows on 4 September 2026 between 16:50 and 19:25 UTC. The later test covered 1,176 windows, using a snapshot taken at an exact second. We did not record the dates of the later sample.

Result

Test Entry Trades Reversals Result per bet
64 windows, favourite at ask 0.90 or more 240 s 51 3 -4.5%
64 windows, same 255 s 59 0 +1.1%
64 windows, spot at least 5 bps from open, in favourite's direction 240 s 13 0 +1.0%
1,176 windows, same 5 bps filter 252 s 14 1 -5.6%

One basis point (bp) is 0.01%. Without the filter the 1,176-window result was negative at every second tested, and entries at 230-245 seconds were negative.

The 64-window "0 reversals" at 255 seconds came from looking ahead: the last two-sided snapshot at or before 255 seconds included information that a real entry could not have had. An earlier figure of "0 of 88" had the same cause.

Where reversals happen

Among candidate trades, reversals were 10% for XRP, 7% for SOL, 3.6% for ETH and 1.5% for BTC. Windows with reversals had volatility of 5.4 basis points per minute against 3.0 without them. By price, a favourite at 0.90-0.96 lost 11% of the time, while one at 0.99 lost 1.7%.

Why the edge disappears

The filter needs to predict the settlement figure, not the Coinbase spot price. We checked: the Coinbase proxy agreed with the Chainlink settlement in only 92% of windows, and in 365 of 1,176 windows the move from the open was under 1 basis point, too small for a proxy to call the direction. That caps how well any spot-based filter can work. The resolution rule also matters: in 12 windows, the outcome matched the "average of the last minute against the open" in 12 of 12, but matched the average over the whole window in 10 of 12.

Limits of this measurement

  • The 64-window test is small, which is why its results fell apart under the stricter method.

  • The 14-trade result of the filter is also very small. One reversal accounts for most of the -5.6%.

  • We did not record the dates of the 1,176-window sample, whether results include the taker fee, or confidence intervals.

  • The spot proxy is Coinbase, not the Chainlink feed used for settlement.

What it means for a trader

We did not find a late-entry edge. The first result looked clean because the snapshot used information from after the entry moment. With a stricter method, the same approach lost on most ways of cutting the data, and reversals concentrated in the more volatile assets, XRP and SOL, and at favourite prices below 0.96. A thin edge of one reversal per 50 to 75 trades cannot be told from noise with a few dozen trades.

Polytrading is an analytics tool. It does not trade for you, hold your keys, or promise returns.