WSS2 groups engineered features by research purpose rather than treating every column as an isolated variable. This makes model tests easier to audit: each run can compare core market state, timing, microstructure, execution quality, regime, strategy context, and quality-control features under the same validation protocol.
| Family | What it captures | Examples | Used for |
|---|---|---|---|
| Core market state | Economic state of the prediction-market contract and underlying spot market at decision time. | moneyness, distance-to-strike, normalized distance, market-implied probability, YES/NO side, prior bid/ask/mid fields | Feature-family toggles in ML grids (see Model Testing Use below) |
| Timing | Where the contract is in its lifecycle and when the decision is made. | time-to-expiry, hour UTC, day of week, entry bucket, late-window flags | Feature-family toggles in ML grids (see Model Testing Use below) |
| Microstructure | Order-book and quote-state features describing tradability and local market quality. | spread, top-of-book depth, depth bands, imbalance, book side presence, quote age, stale-book flags, alignment gap | Feature-family toggles in ML grids (see Model Testing Use below) |
| Execution quality | Features describing whether a paper signal could plausibly be executed at q=10 under conservative assumptions. | q10 availability, vwap, simulated fill cost, depth consumed, slippage proxy, limit-fill feasibility | Feature-family toggles in ML grids (see Model Testing Use below) |
| Regime | Coarser state buckets used to test whether strategy performance changes across market conditions. | near/far strike bucket, tight/wide spread bucket, high/low depth bucket, liquidity regime, timing regime, implied-probability regime, trend/chop/volatility regime if safely prior | Feature-family toggles in ML grids (see Model Testing Use below) |
| Strategy context | Features describing which public strategy family generated the candidate and how strategy families interact. | base strategy family, DOVE/PACE/SIEVE/LATE/FTE/DART indicator, strategy side, strategy agreement/disagreement state if known before selection | Feature-family toggles in ML grids (see Model Testing Use below) |
| Quality controls | Flags used to prevent invalid or low-quality observations from entering model training or scoring. | missing book, negative spread, settlement-edge flag, ML eligibility flag, data freshness flag, missing side flags | Feature-family toggles in ML grids (see Model Testing Use below) |
Outcome, realized PnL, settlement, and post-entry information are excluded from model inputs. Feature availability is checked against decision time before training.
Feature families can be toggled in ML grids to test whether a class of information adds value beyond simpler baselines. For example, a model can be trained on core market state only, then retested with timing, microstructure, execution, or strategy-context features added.
Latest generated: 2026-07-07T01:14:45Z.
Main dashboard · Research Watchlist