Reverse-engineered where the high-edge wallets (edge_z > 2) on pm_short_term_trader_accuracy actually make their money. The naive “Binance leads, book lags, exploit the lag” thesis from short-term-accuracy and lag-probe failed M0 backtesting at 46% win rate. The real edge is not latency arb. This note replaces the strategy thesis.
Primary case: 0x75cc3b63a2f2423085e10706c78b494017b93ce1. Cross-validated on the other 8 wallets in the cohort.
headline
The book’s own structural under-pricing of residual volatility late in a 5-minute window is the edge — especially on the cheap-tail side. The book treats nearly-dead outcomes as deader than they are. BTC-vs-book lag is real (see lag-probe: p50 300 ms) but it isn’t what these wallets are harvesting.
what 0x75cc3b actually does — two lanes
Both lanes operate only in the last third of a 5-minute window (time_in_window > 0.66), 5m markets only, cross-asset (BTC + ETH + SOL + XRP).
Lane 1 — late-window longshot reversion
Buy at best-ask ≤ $0.30 when an outcome looks “almost dead.” The book over-prices the deadness; residual underlying volatility flips outcomes more often than the 5%-of-implied book suggests.
Dollar-weighted ROI by entry price band:
| price band | ROI |
|---|---|
| $0.00–0.10 | +132.8% |
| $0.10–0.30 | +25.0% |
| $0.30–0.50 | (decreasing) |
| … | … |
| $0.90–1.00 | ~+1% |
Monotonic from cheap to expensive.
The phase × price heatmap is the killer:
| phase | $0.00–0.10 ROI |
|---|---|
| late-window | +222% |
| early-window | −68% |
Same band, same wallet, opposite sign. Timing is half the edge.
Lane 2 — late-window momentum-chase
When the book has bid an outcome up ≥ 2¢ in the last 30 s and we’re in the last third of the window, take the offer (cap ~$0.85).
- Hit rate: 73.4%
- Edge: +4.7¢
The book lags its own recent moves; consume the slow liquidity.
what was wrong about 0x75cc3b before
| was believed | actually |
|---|---|
| 3,397 markets | 28,696 BUYs across 43,734 trades. pm_short_term_trader_accuracy.n_markets counts distinct outcomes touched, not trades. |
| BTC only | BTC + ETH + SOL + XRP updown markets |
| Median entry 15% into window | Median 41% for 5m, 44% for 15m — fairly uniform; edge is concentrated late, but trades are not |
| One strategy | At least two lanes (longshot + momentum), 5m-specific |
generalizes — partially
| wallet | pattern | strength |
|---|---|---|
0xd189664c | PUREST: 613 late-window 5m longshot trades at 9.4K PnL on $12K invested** | strongest single-wallet validation |
0x8c901f67, 0x20d2309c, 0xeebde7a0 | same pattern, weaker (0xeebde7a0 looks like spread-capturing MM) | confirms lane 1 |
0x7523cafc, 0xb27bc932 | DIFFERENT strategy: favorites-only in 15m markets | at least 2 distinct edge-bearing strategies coexist in the cohort |
what to avoid (per the data)
- 15m BUY Up at any price — loses for every asset
- 5m early-window low prices (early × $0–0.10 = −68% ROI)
- 5m mid-window deep favorites — basically zero edge with capital tied up
strategy v2 implications
Redesign the bot around:
- Signal source: book state + time-in-window. NOT Binance vs book.
- Universe: cross-asset by default (BTC + ETH + SOL + XRP), 5m markets primarily.
- Lanes: longshot reversion (price ≤ 0.85).
- Infrastructure: less about Binance feed microseconds, more about staying connected to a fast Polymarket book feed and computing late-window reversal probabilities. The eu-west-2 colocation argument from short-term-accuracy still holds — we still need to be fast on the CLOB — but the signal is endogenous to the book, not the BTC tape.
This deprecates the M0 model from the lag-probe capture as the primary strategy. The lag measurement (p50 300 ms book-lags-Binance) is still a true fact about the market; it just isn’t where this cohort’s money comes from.
caveats
- No external price tape used in this analysis — momentum was inferred from Polymarket’s own book activity. Whether “book lags its own activity” vs “book lags Chainlink/Binance” needs a synced feed to disambiguate.
- Order-book depth/spread unobserved; only fills visible. Maker vs taker not distinguished —
0xeebde7a0’s 20+ trades/outcome suggests maker behavior (spread capture) rather than pure direction. pm_tradesis restricted to ~119 watched wallets; “momentum” computed over that subset is a proxy, not the full book.- 7-day observation window — niche cells (e.g. late deep-longshot, n=478) carry ~1–2¢ edge uncertainty per cell.
related
- short-term-accuracy — the earlier (now-corrected) latency-arb framing
- lag-probe — measured the lag we thought was the edge; still informs the execution infra need
- copy-trade-algorithm — sibling algo, all-category; same survivorship-bias caveats apply
- polymarket-fetch — overview / schema / CLI