Winners Harvest Mechanical Structure, Not Predictions
After dissecting 43 purged-CV-robust strategies across four venues, every winner mapped to a mechanical quirk. Settlement-window convergence, maker rebates, oracle latency, cross-asset lead-lag. Pure directional bets won zero seats. The edge-class distribution and what it means for strategy design.
The setup
We ran the same fleet — about 2,400 trading personas across six universes — through a regenerated, contamination-audited corpus (how the audit was done). After purged combinatorial cross-validation, 43 strategies were robust. Then we looked at what they were.
The result was the kind of distribution that, in retrospect, is obvious — and which essentially nobody publishes about, because it's not the kind of thing you can sell on Twitter:
| Edge class | Share of winners | Example |
|---|---|---|
| Maker / LP / incentive (E07–E12) | ~35% | Limitless 100% maker rebate, Kalshi liquidity incentive program |
| Structural arbitrage (E02–E06) | ~20% | Pair-cost on neg-risk legs, calendar arb |
| Microstructure / oracle-lag (E13–E17) | ~20% | NWS gridpoint weather oracle lag, settlement-window convergence |
| Statistical MR (E18–E20) | ~10% | OU mean reversion, round-number fade |
| Event / regime (E21–E24) | ~10% | Post-debate drift, macro-event anticipation |
| Directional (E01) | ~0% | Pure outcome prediction — zero robust winners |
Read the bottom row carefully. Zero of 43 robust winners are directional bets. Not a small fraction. Zero. The IRL experiment from the previous post independently confirmed why: directional alpha on calibrated markets is unlearnable from decision-time features.
Five winners, dissected
These are public versions of strategies in the proven-positive set. Names anonymized to persona-class.
1. settlement_window_scalper_t1 — +$1,292 paper P&L on Kalshi
Edge class: E16 — resolution-boundary convergence arbitrage.
Why it works: on a binary market with a known resolution timestamp, the contract price must
converge to either 0 or 1 by resolution_ts. In the final ~5–15 minutes before resolution, the market's
information signal-to-noise is dominated by the convergence trajectory rather than any new information. Sizing a small
book that fades away from {0, 1} into {0, 1} extracts the convergence premium without taking a directional bet.
The persona doesn't predict the outcome. It predicts that the price will be closer to 0 or 1 at T+15 than at T.
2. offhours_lp_t1 — +$1,211 on Kalshi
Edge class: E09 — regime-gated maker harvest. Why it works: high-frequency competitors retire during off-hours (UTC nights, weekend pre-market). The book's spread widens, adverse selection from informed flow drops to near zero, and the maker rebate becomes the expected value of the trade. Restricting the persona to a low-competition regime window converts a structurally money-losing market-maker into a money-making one.
The edge is the absence of competition, not the presence of insight.
3. volume_incentive_optimizer_t1_flip — +$1,009 on Kalshi
Edge class: E11 — fee-schedule cross-tier accounting. Why it works: Kalshi's liquidity incentive program (LIP) pays a per-volume bonus that crosses tier boundaries. Trades that cross a tier boundary are worth more than trades that don't, because the bonus applies to the next-tier rate on all qualifying volume. A persona that times its volume to land on the correct side of the boundary at month-end can extract a few hundred basis points of pure fee revenue.
This is not a trade. It's an accounting strategy that uses trades as the mechanism.
4. climate_regime_t2 — +$503, battery-clean on Kalshi
Edge class: E15 — oracle-lag arbitrage. Why it works: Kalshi's daily-temperature contracts resolve on the National Weather Service gridpoint observation. The NWS API publishes updates at fixed intervals; our feeder polls those intervals; the market reprices on a slower cadence than the source publishes. In the gap between "the truth has changed" and "the market knows the truth has changed," there is a small but persistent edge.
The market resolves on a specific data source. Reading that source faster than the market reprices is the entire edge.
5. f698_xasset_btc_futures_alpha_leak_t2 — +$640 on Kalshi
Edge class: E23 — cross-asset lead-lag. Why it works: Kalshi's BTC-price prediction contracts are denominated against the BTC/USD price. The CME and Coinbase BTC orderbooks lead Kalshi by a measurable, latency-bounded interval. When BTC ticks on CME but Kalshi hasn't repriced, there's a brief arbitrage window. The persona doesn't predict BTC's direction — it predicts that Kalshi will, within the next second, catch up to where CME already is.
The directional risk is hedged by construction; the latency lag is the edge.
The taxonomy
We built a 24-class edge taxonomy ("E01–E24") to map the universe-by-edge-type coverage of our fleet. The classes cluster into six families:
- Directional (E01): bet the outcome. Zero robust winners. Stop doing this on calibrated markets.
- Structural arbitrage (E02–E06): pair-cost, neg-risk, cross-venue, combinatorial, calendar. Reliable but capacity-limited.
- Maker / LP / spread / rebate (E07–E12): the fee schedule pays you to provide liquidity. Highest robust-winner share.
- Microstructure (E13–E17): OFI, queue position, oracle lag, settlement window, VPIN. Most technically demanding, real edge when implemented carefully.
- Statistical MR (E18–E20): OU, GARCH, overreaction. Modest, regime-dependent.
- Event / regime (E21–E24): earnings, debates, lead-lag, liquidity vacuums.
The per-venue map
Where you trade matters more than what you trade. Each venue's edge profile is structural, not random — it's a function of fee schedule, counterparty composition, oracle architecture, and incentive program.
| Venue | Best-fit edge classes | Why |
|---|---|---|
| Polymarket-US | E04 (cross-venue), E07 (maker), E15 (oracle-lag) | 0.20% maker rebate, UMA oracle latency, neg-risk structure |
| Kalshi | E11 (fee-tier), E15 (oracle-lag), E16 (settlement-window) | LIP fee structure, NWS/FRED feeders, CFTC-licensed |
| Limitless | E07 (maker), E12 (incentive) | 100% maker rebate, point/airdrop program |
| IBKR ForecastTrader | E04 (cross-venue arb), E15 (oracle-lag) | Unifies Kalshi+CME+ForecastEx in one account; ~3.14% APY on collateral |
| Coinbase Derivatives | E22 (funding-rate) | 0.02%/contract, 10x leverage, perpetual funding arb |
The same persona ported across venues often goes from money-making to money-losing depending on fee schedule alone. Edge is not a property of the strategy. It's a property of the strategy × venue × regime triple.
Implications for strategy design
1. Pick the edge class first, the strategy second.
Walk-forward research that starts from "let's find a signal" almost always converges on a directional bet, because that's where the easiest features live. Walk-forward research that starts from "we want to harvest E15 oracle-lag on Kalshi weather contracts" produces a strategy with a structural reason to be profitable. The framing matters.
2. Map your venue's structural quirks.
Every venue has 3–8 structural quirks (fee tier, rebate, oracle architecture, resolution mechanism, incentive program). Most of them are documented in the venue's API docs and ignored by quants chasing directional alpha. Read them.
3. Incentive programs end. Plan accordingly.
Kalshi's LIP expires 2026-09-01. Limitless points may not become airdrop. PM-US's maker rebate could be cut at any time. Strategies that depend on incentive programs have a half-life. Size accordingly, and have a successor strategy ready when the program ends.
4. Capacity is the real ceiling on structural edge.
Settlement-window convergence on Kalshi has maybe $50–500K of capacity before your fills become the entire orderbook. Maker-rebate harvest on Limitless has more, but is bounded by the platform's total daily volume. Structural edge scales sub-linearly. The +$12,814 the kalshi proven-set produces is not 33× that on a 33× larger book; it's more like 5–8×.
5. The proven-set is a starting point, not a finish line.
Out of the 43 strategies that survived purged-CV, 0 passed the full 26-method validation battery. Five are "battery-clean and forward-paper positive" but pending the 14-day forward-paper SPAN gate. That's the honest number, and it's a much smaller number than "43 winners." The lesson is that purged-CV is necessary but nowhere near sufficient.
What to do with this
- If you're building a quant strategy on a prediction market, your first question should be "which of E02–E24 am I harvesting?" If the answer is "E01" (directional), reconsider.
- If you're already running directional strategies that look profitable in backtest, you have ~80% likelihood of a contamination artifact — see the corpus audit post.
- If you want the full taxonomy + the per-venue coverage matrix, it's in the white paper.
- If you want us to map your strategy onto the taxonomy and run the validation battery, that's Validation-as-a-Service.
Honest disclosure. These posts come from an internal trading-research program ($5K experimental capital, paper-mode preserved). Results are reported as measured; none of this is investment advice. Where a method or finding has caveats, we name them in-line — that is the whole point of this series.