himanshuj16/algo-trading-skills/adverse-selection-measurement-for-passive-orders

When to Use

Invoke this skill to evaluate the execution quality of passive / liquidity- providing algorithms — market makers, passive limit strategies, and any resting-order logic. If your fills consistently happen right before the market moves against you (you buy, price drops; you sell, price rises), you are suffering adverse selection: informed flow is picking off your resting quotes, and you are writing a free option to the informed (Glosten & Milgrom, 1985).

The skill produces a MarkoutEngine that computes, for each fill and each forward horizon, the basis-point price drift from the fill (or from the fill- time mid) to the future mid, then aggregates into a per-horizon markout curve with distribution statistics. A persistently negative curve = toxic flow.

When NOT to Use

• Active / aggressing orders (market orders, crossing sweeps). Adverse

selection is a passive phenomenon — you were resting and got picked off. Active orders pay the spread up front; measure that with execution-slippage- attribution-timing-vs-sizing or implementation-shortfall-minimization.

• Latency-arbitrage diagnostics. A sharp negative markout in the first

1–10 ms is a symptom of stale-quote latency arbitrage against you, but the fix lives in the feed-handler/cancellation path (tick-to-trade-latency- measurement), not here. Use this skill to detect, then route to the latency skill to remediate.

• Alpha / signal research. Markouts measure execution friction, not

predictive power. For signal strength use backtest-reporting-standardized- tearsheet or factor research skills.

• Markets without a clean mid (some OTC, illiquid single-name options with

wide stale quotes). The mid-to-mid markout is only as honest as the mid; a stale or gappy mid produces nonsense markouts. Use multi-source-price- reconciliation-tie-breaking to fix the mid first.

• Tick-by-tick attribution of a handful of fills. With < ~30 fills per

horizon the mean is noisy; report the distribution (median, IQR) and do not over-interpret the sign of the mean.

Prerequisites

• Python 3.10+, numpy.
• A ledger of passive fills — each a `(trade_id, timestamp, side, fill_price,

quantity)`. Active/aggressing fills must be filtered out upstream; this engine assumes every fill it receives is passive.

• A high-resolution mid-price series (timestamp, mid) covering every fill

timestamp and max(horizons) seconds forward. Timestamps must be strictly ascending, finite, with positive mids (validated at evaluation time).

• Clock alignment between the fill ledger and the market-data series —

same epoch, same time base. See clock-skew-correction-for-tick-timestamps.

Workflow

1. Filter to passive fills. Exclude market/aggressing orders upstream;

adverse selection only applies to resting liquidity.

1. Configure horizons and basis:

```python from adverse_selection_measurement_for_passive_orders import ( MarkoutEngine, MarkoutConfig, PassiveFill, )

config = MarkoutConfig( horizons_sec=[0.1, 1.0, 5.0, 60.0], # 100ms, 1s, 5s, 1m markout_basis="fill_to_mid", # or "arrival_to_mid" (mid-to-mid) quantity_weighted=True, # share-weighted MEAN only require_asof_mid=True, # no-lookahead guard, both bases max_mid_staleness_sec=None, # set it; see step 2b ) engine = MarkoutEngine(config) ```

Set the staleness bound deliberately. max_mid_staleness_sec is None (off) by default because no universal value exists — the engine will not invent one. Derive it from the instrument's own quote-update cadence: a bound well above the typical inter-quote gap catches only session gaps, halts and dead feeds; a bound near that gap discards live data. Left unset, a markout can be measured against a mid from before a halt and still look like clean evidence.

1. Evaluate:

``python report = engine.evaluate_fills(fills, market_timestamps, market_mids) ``

1. Check the data verdict before the toxicity verdict.

• report.has_sufficient_data / report.evaluable_horizons — how many

horizons actually produced a markout. If this is 0, `is_toxic=False` means "not measured", not "healthy" — the message reads INSUFFICIENT DATA. Never gate on is_toxic without checking this first.

• report.missing_pre_fill — fills skipped by the no-lookahead guard.
• report.stale_asof_mid — fills skipped because the fill-time mid was

older than max_mid_staleness_sec.

• report.stats[h].truncated — fills dropped at horizon h because the

market data did not cover fill_ts + h.

• report.stats[h].stale — fills dropped at h because the prevailing mid

there was over-age.

Then read the curve:

• report.average_markouts_bps[h] — mean markout per horizon (backward-compat).
• report.stats[h] — count, mean, median, p25, p75, std, truncated, stale.

Only mean_bps honours quantity_weighted; the median, quartiles and std are always unweighted order statistics.

• report.is_toxic — True if a majority of the evaluable horizons have a

negative mean. Horizons with no data are excluded entirely rather than counted as healthy.

• report.toxicity_ratio — negative share of the evaluable horizons.

1. Diagnose the shape (see Decision Points):

• Sharp negative in the first 100 ms → stale-quote / latency arbitrage.
• Gradual negative over seconds–minutes → directional adverse selection

(your alpha is wrong or you are on the wrong side of informed flow).

• Positive at short horizons, negative later → you capture the spread but

bleed to informed flow over the holding period.

Decision Points

Common Pitfalls