The Strategy Factory

An automated pipeline that transforms hypotheses into validated prediction market strategies. Seven stages, three statistical gates, zero tolerance for overfitting.

Why PM backtesting is different

In traditional quant trading, you backtest against 20 years of continuous price data for the same instrument. In prediction markets, every market is unique — born, lives briefly, resolves to 0 or 1, and dies. You can't backtest "a strategy on ECB March 2026." You backtest "a pattern across all ECB rate decisions as a class."

The Strategy Factory solves this by shifting the unit of analysis from instruments to event-type categories.

Seven stages, four gates

Every strategy flows through the same pipeline. At each gate, it either passes or is rejected. No exceptions, no shortcuts, no re-tuning after failure.

1
Hypothesis
Pattern identified
2
Formalize
14-field spec
3
Train
60% sample
GATE 1
4
Validate
20% held-out
GATE 2
5
Out-of-Sample
20% one-shot
GATE 3
6
Paper Trade
30+ events
GATE 4
7
Live
2% capital
The Factory in one sentence

An automated pipeline that transforms hypotheses into validated strategies, deploys the survivors with risk-controlled capital, monitors them against statistical benchmarks, and retires them when edge decays — while simultaneously developing replacements.

Three-sample testing for event markets

The classical mechanical trading system framework — Training, Validation, Out-of-Sample — adapted for prediction markets where samples are split across events, not time periods.

Training — 60%
Validation — 20%
OOS — 20%
Full optimization permitted. Overfitting expected.
NO parameter adjustment. Where overfitting dies.
One-shot. If it fails, the strategy is dead.
60%
Training
Parameter optimization on majority of historical events. Sharpe > 1.0, Win Rate > 55%, EV > 0. Profitable at +25% fee assumption.
20%
Validation
Held-out events. NO tuning allowed. Degradation must be < 50% from training. If it fails here: reject or return to hypothesis.
20%
Out-of-Sample
One-shot final test with realistic execution model. Profitable at +50% fee assumption. Fail = dead. No second chances. No re-tuning.

Why this works for small samples

Traditional quant backtesting uses thousands of data points. PM event categories have tens to hundreds. The Strategy Factory compensates with: conservative parameter selection (wide robustness windows), mandatory paper trading before live deployment, small initial position sizes (2% of capital), rapid feedback loops, and — critically — the Structural Event Clustering that pools structurally similar events across categories to increase effective sample sizes.

Two layers, Polymarket-anchored

The Strategy Factory runs two complementary layers: directional alpha (single-leg trade on Polymarket, edge from a trading rule or trusted external source) and arbitrage (both legs tradable, risk-free pair-trade).

Single-leg position on Polymarket — edge from rules or source-of-truth

In directional-alpha trades only one leg is tradable: the Polymarket position. Edge comes either from classical trading patterns or from an external source whose authority we accept.

1a — 51-Pattern Trading Rules

Adaptation of classical quant patterns (mean reversion, volume spikes, trend following) onto prediction-market mechanics. Pipeline pattern is validated step by step.

SF-PAT-001In progress
51-Pattern Pipeline
Cross-domain · Polymarket
51 patterns
per pattern validated
seq. roll-out
Adaptation of classical trading rules from equities/futures onto PM mechanics. Detail at Directional Alpha.

1b — Source of Truth vs Polymarket mispricing

Treat a trusted external source as “truth” — when Polymarket diverges by more than the threshold, trade the Polymarket leg. The source itself is usually not tradable for EU users; that's why this is single-leg, not arbitrage.

SF-STR-001Validation
Pinnacle Sharp Reference (sport season)
Sport season outright · Polymarket vs Pinnacle
~62% win
3-7pp edge
50-100/yr
Source of truth: Pinnacle vig-adjusted (~2 % margin → 1.0 normalised) as the sharp-money benchmark. Polymarket season markets (Champions League qualification, relegation, league winner, top scorer) vs Pinnacle outrights.
SF-STR-002Validation
FOMC implied vs Polymarket
US rates · Polymarket vs CME ZQ
~60% win
3-8pp edge
8-12/yr
Source of truth: Fed Funds Futures (ZQ via EODHD) → CME methodology implied. Live in the Mispricing Lab.
SF-STR-003Validation
Macro consensus spread
US macro · Polymarket vs Finnhub
~58% win
2-6pp edge
12+/macro event
Source of truth: Finnhub Economic Calendar consensus for CPI/PPI/NFP/GDP. Cleveland Fed nowcast and Bloomberg consensus as cross-checks planned.
SF-STR-004Q3 2026
ECB rates vs OIS curve
EU monetary policy · Polymarket vs OIS
~62% win
3-5pp edge
8-12/yr
Source of truth: €STR OIS forwards for ECB meetings. EU OIS adapter pending — Q3 2026.
SF-STR-005Hypothesis
Bundesliga info edge
Sport season · Polymarket + DE/IT/FR news
~60% win
3-6pp edge
30-60/yr
Hybrid: Pinnacle as anchor plus a language edge — kicker.de, Transfermarkt, Gazzetta, L'Équipe propagate injury and transfer news 2-6 h ahead of English sources.
SF-STR-006Hypothesis
EU coalition / regulatory
EU politics · Polymarket
~65% win
5-15pp edge
5-10/yr
Own models: coalition arithmetic, Brandmauer logic, EU institutional knowledge (MiCA, DSA, ECJ timelines). US bots model “highest poll = winner”.

Both legs tradable — cross-platform pair trade

In arbitrage both platforms are tradable — the trade is mathematically risk-free as soon as the spread exceeds the sum of all costs. V1: Polymarket vs other prediction-market and sports-exchange platforms. Most second-platform adapters Q4 2026.

SF-ARB-001Q4 2026
Polymarket vs Kalshi
US markets · Polymarket + Kalshi (read-only)
~80% win
2-4pp edge
20-40/yr
Settlement-rule-verified identical events with ≥ 5 pp price difference. Kalshi read-only API hookup pending.
SF-ARB-002Q4 2026
Polymarket vs Smarkets (sport season)
EU sport season · Polymarket + Smarkets
~70% win
3-6pp edge
30-50/yr
UK-tradable sportsbook exchange. Season outrights (champion, top-4, relegation) on both — spread on identical outcomes.
SF-ARB-003Q4 2026
Polymarket vs Manifold
Cross-platform · Polymarket + Manifold
~75% win
5-15pp edge
10-25/yr
Manifold liquidity is small but the API is open. Useful for long-tail markets where Polymarket liquidity is thin.
SF-ARB-004Q4 2026
Polymarket vs Betfair (match outright)
UK/EU sport · Polymarket + Betfair
~65% win
2-4pp edge
20-40/yr
Betfair Exchange for season outrights and top-tournament events. Different market structure (1X2 vs YES/NO) — mapping required.

Anti-overfitting rules

Overfitting is the primary risk in PM backtesting due to small sample sizes. These rules are system-enforced, not suggestions.

No re-tuning after validation failureReturn to hypothesis stage, not training. New specification required.
Out-of-sample is one-shotIf it fails, the strategy is dead. No second chances, regardless of backtest strength.
Fee sensitivity mandatoryEvery strategy must be profitable at +25% fees (validation) and +50% fees (OOS).
Parameter robustness windowOptimal parameters must remain profitable at ±15%. Cliff-edge parameters = overfitting.
Paper trading required30 events or 8 weeks minimum. No exceptions.
Max 15 live strategiesForces prioritization. Prevents capital dilution across too many marginal strategies.

The Factory is operational

v1.0 running since February 2026. First strategies in the pipeline. The methodology is proven — now being productized as infrastructure.

See the Innovation Layer → Trading Tools →