Research/Nov 10, 2026

Copying profitable wallets with a delay: no edge after 5 minutes

Copying a profitable wallet's taker buy 1 minute late returned +0.8% per $1 (standard error 3.0 points); 5 minutes late, -4.8% (3.3); 30 minutes late, -15.8% (3.1). Wallets with losing records did about as badly.

What we asked

Copy-trading tools assume that the purchases of a historically profitable wallet carry information, and that a follower can capture some of it by buying the same outcome shortly afterwards. We tested whether that survives a realistic delay. A taker is a trader who buys against an order already resting in the book, which makes the trade visible immediately. We took taker buys by "leader" wallets and asked what a follower would have earned by copying them 1, 5 or 30 minutes later.

This is a paper simulation from public data.

Data

  • Cutoff. Wallets were classified using data up to 15 September 2026 00:00 UTC. Trades after the cutoff, through 14 October 21:00 UTC (29.9 days), were used only for measurement.
  • Leaders. From a random sample of 5,160 wallets, a leader candidate had at least $500 of total profit and loss (P&L) on 15 September and positive P&L over the preceding 60 days. After filters for trade count, taker share and trading frequency (meant to exclude market-making bots), we took 150 leaders. The 150 comparison wallets ("outsiders") had negative P&L on 15 September and passed the same trade-feed filters.
  • Signals. A signal is a taker buy after the cutoff in a binary market that has closed and resolved strictly to 1 or 0 and has at least $20,000 of lifetime volume, capped at the latest 100 signals per wallet. There were 6,275 leader signals from 127 wallets and 3,045 outsider signals from 76 wallets.
  • Copy price. The last price in the CLOB price history at or before the leader's trade plus the delay, plus 1 cent of slippage. The position is held to resolution. Expected value (EV) per $1 is payout divided by price, minus 1.
  • Uncertainty. Figures in parentheses are bootstrap standard errors, in percentage points, resampling wallets rather than trades.

Result

Group Leader's own price Copy +1 min Copy +5 min Copy +30 min
Leaders, all prices (n = 6,275) +17.2% (16.3) +0.8% (3.0) -4.8% (3.3) -15.8% (3.1)
Outsiders, all prices (n = 3,045) -4.5% (3.1) -5.3% (3.2) -9.0% (3.1) -16.3% (3.5)

"Leader's own price" is the same outcome valued at the price in the leader's trade, with no delay and no slippage. It is not the leader's actual P&L, because it assumes every buy is held to resolution.

The leaders-minus-outsiders gap, with a 95% bootstrap interval, was +6.1% [-2.0; +14.6] at 1 minute, +4.3% [-4.0; +13.4] at 5 minutes and +0.5% [-8.9; +8.9] at 30 minutes. Every interval includes zero.

The +17.2% at the leader's own price comes mostly from outcomes priced below 0.2, where it was +149.3% with a standard error of 170.0 points. A handful of wins decides that band.

Why the edge disappears

Our report does not isolate a single cause, but the pieces are visible. The copy pays 1 cent of slippage on top of the history price. The median price move after a leader's trade was about -0.5 cents at every delay. The mean move for leaders was +0.37 cents at 5 minutes and +0.91 cents at 30 minutes, meaning the price had on average moved against a late buyer. In the 0.2-0.5 band the mean move at 30 minutes was +1.48 cents, and the copy lost 17.9% (5.5).

Outsiders' results worsen with delay in the same way, from -5.3% to -16.3%.

Limits of this measurement

  • The price history is a series of prices, not an order book. It has no depth or spread, so real execution at the ask could have been worse. The copy's size was not modelled: a real copy of the same size would also consume depth.
  • About 30 days and 127 leaders with signals is a short, thin sample. The bootstrap resamples wallets and does not capture correlation between wallets trading the same events.
  • Only closed and resolved markets with at least $20,000 of volume were included, which biases toward fast-closing markets.
  • The wallet list comes from the trade feed around 14 October, so wallets that stopped trading are absent. The "not a market maker" filter relies on taker share and frequency only. The P&L series probably includes open positions at current prices; this was not verified.

What it means for a trader

In this sample, buying a profitable wallet's taker purchase 5 or 30 minutes later returned a loss, and 1 minute later returned a result statistically indistinguishable from zero (+0.8%, standard error 3.0). The comparison with losing wallets did not separate the two groups. The sample is too small and too short to claim that copying cannot work, but we found no delay at which it measurably did.

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