Hey — quick rundown on the idea you texted me. I actually went and built it. Spoiler: doesn't work, but the why is more interesting than the verdict.
I built a tool that tests your "copy the top wallets on Polymarket" idea against ~2 years of real historical data, without ever risking a dollar. The short answer: the idea has a real kernel of truth, but it loses money once you account for realistic trading costs. A clean "no, don't trade this." Details below — and the why it doesn't work is the interesting part.
"Polymarket has a public leaderboard of the most profitable wallets. If we copy their bets right after they make them, do we make money? Pelosi-tracker, but for prediction markets."
A perfectly reasonable thing to ask. There's a long history of "copy the smart money" working in some contexts (Congressional traders, hedge fund 13F filings, etc.). Worth checking whether it works here before ever putting real money down.
- Polymarket is a prediction market — people bet on yes/no questions ("Will X happen by Y?"). Each yes/no share pays $1 if it wins, $0 if it loses, and trades anywhere between those before the market ends ("resolves").
- The leaderboard ranks wallets by realized profit. Anyone can see it.
- The catch: when you actually look at the top of the leaderboard, most of those wallets aren't smart bettors — they're bots that quote both sides of every market and collect the spread (called market-makers and arbitrageurs). They make money on volume and microstructure, not by being right about outcomes. Copying a bot is pointless — by the time you've entered, the tiny mispricing they exploited is already gone.
- So the real question becomes: among the top wallets, can we tell the actual directional bettors ("smart money") apart from the bots, and is that subset profitable to copy?
That's what I built and tested.
A local research tool (~4k lines of TypeScript). Nothing live, nothing risky — purely analysis:
-
Data pull. Downloads everything from Polymarket's public APIs into a local database: who's on the leaderboard, every trade those wallets have ever made, every market they traded, and how those markets resolved. (Top wallets do hundreds of thousands of trades each, so this took some engineering.)
-
Wallet classifier. A set of rules that looks at each wallet's behavior and labels it as one of:
- Smart money — patient directional bettor; holds positions for hours to weeks, wins more than half the time on real bets.
- Market-maker / arbitrageur — bot; flips constantly, holds both sides, profits from microstructure.
- Unknown — not enough history to judge.
-
Backtester. Simulates what would have happened if we had been copying the smart-money wallets, week by week, over the past two years. Crucially it does this walk-forward, which means: at each past date, the decision of who to copy is made using only data that existed before that date. No cheating by knowing who turned out to be a winner. This is the single most important honesty feature — every "this strategy worked in backtest!" story you've ever heard is usually broken because they cheated on this.
-
Realistic costs. Every simulated bet pays:
- A delay — you don't see the smart wallet's bet instantly, so you enter 15 minutes later at a worse price.
- Slippage — the act of buying a bunch of shares pushes the price against you. Bigger bet = worse slippage.
-
Robustness check. Re-ran the whole backtest 108 times across different combinations of: delay length, position size, watchlist size, and how strict the classifier is. This is how you tell "real signal" from "I got lucky on one config."
- 50 wallets from the current Polymarket profit leaderboard.
- Their full history: ~1.9 million trades and ~32,000 markets going back to mid-2023.
- The classifier sorted them into: 10 smart-money · 16 bots · 24 unknown. (So 60%+ of the top leaderboard is bots — your intuition was right that there's a lot of noise up there.)
- Backtested copying the 10 smart-money wallets every week from June 2024 → May 2026 (~23 months).
Before I looked at any numbers, I wrote down what "this works" would have to look like:
- Make money after costs (positive return after slippage).
- Be worth the risk (Sharpe ratio > 0 — Sharpe is just "how much return per unit of bumpiness"; negative means you lost money on a risk-adjusted basis).
- Hold up under stress-testing — be profitable across most of the 108 parameter combos, not just one cherry-picked corner.
Setting this in advance matters because once you start staring at numbers, it's very easy to talk yourself into "well it's almost positive if I just tweak…" That's how people lie to themselves with data.
| Bets copied | 1,402 over 23 months |
| Money put to work | $1.4M total |
| Bottom line | −$8,706 (−0.62%) |
| Sharpe | −0.13 (worse than just holding cash) |
| Worst peak-to-trough drop along the way | 3.9% |
| Win rate | 54.8% ← remember this |
Across the whole grid:
- Best corner of the entire grid: +1.5% over 23 months — about +0.8% per year — and only when using the smallest $250 bets where slippage hurts least. Below treasury bond rates. Not a strategy.
- Most corners are negative. As soon as bets get to $1k or larger, or the delay stretches past 30 minutes, every single config loses money.
- 0 of 4 stress-test dimensions held up. The pre-set "≥3 of 4" bar was not even close to being met.
-
Win rate > 50% but losing money. Why? The classifier did pick wallets that are right more than half the time (54.8%). They just lose bigger when they're wrong. Being directionally correct on the majority of bets is not the same as being profitable to copy if your wins are small and your losses are large. This is the central insight: a profitable bettor and a profitable wallet-to-copy are not the same thing.
-
There's a real edge in sports — it just isn't big enough.
- Sports bets: +$40,409 profit (919 bets) → genuine signal
- Everything else: −$50,273 (456 bets) → wipes out the sports edge and then some
- The "smart money" we're copying is genuinely good at sports markets — and apparently overconfident on everything else.
-
The cost wall is what kills it. Every time you increase position size or increase the delay, the result gets worse. The signal is too thin to absorb realistic trading friction. This is a structural problem, not something you fix by trying harder.
-
The classifier matters. Copying every wallet on the leaderboard (no filter) loses ~3%. Copying only the smart-money-filtered subset is much better (closer to break-even). So filtering out bots does add value — it's just that the filtered subset still can't beat the cost wall.
- The 10 "smart money" wallets came from today's top-50 leaderboard. That means we're already starting with wallets we know became successful — survivorship bias. If anything, this should make the backtest look better than reality. It still came out negative.
- The watchlist was usually tiny — 0 to 4 wallets qualified at any given week (out of 10 in the pool, narrowed further by the walk-forward rules). High variance.
- The Polymarket leaderboard structurally caps the universe at ~50 wallets per window, so we can't easily widen the candidate pool without changing the sourcing approach.
- Sourcing wallets outside the top-50 leaderboard. Catching big directional bettors who don't rank by pure profit. Would probably make the verdict even more clearly "no," but worth confirming.
- Smarter bet sizing (weighting by wallet quality instead of flat $1k each). Might tighten the edge slightly; unlikely to break the cost wall.
- A sports-only version. Sports was the only category with a real edge in the data. There might be something narrower there. Still constrained by costs, but the per-bet edge looked bigger.
You asked a smart question. There's a real instinct behind it — that informed bettors on prediction markets should leave some alpha for a copier to capture, the way Congressional traders apparently do. The data says: not enough, after costs, for this universe.
A clean "no" is genuinely as valuable as a "yes" would have been. We answered the question cheaply, in code, before any money was at risk — instead of finding out the slow way. The whole point of doing it this way was so we could quit without it costing us anything.
If any of the "didn't test" angles is interesting to you, ping me — happy to run the sports-only version or the wider-universe pass.
— Travis