Library/Market Microstructure/Liquidity and Spreads

D11-F02-A04 / Complete engineering topic

Roll Spread Estimator

A production-minded guide to Roll Spread Estimator.

D11 · MARKET MICROSTRUCTURE
D11-F02-A04Canonical / Tested / Open
D11 / D11-F02

Can bid–ask bounce reveal an implicit spread when quote data are unavailable? The answer matters because a visible quote, a signed execution, a later midpoint, and a low-frequency proxy describe different objects. This tutorial gives practitioners a usable measurement and gives builders the data, clock, and validation contract needed to reproduce it.

Start with the decision

Alternating bid- and ask-side prints can induce negative serial covariance even when efficient-price changes are uncorrelated. The canonical package selects Sample covariance of adjacent price changes with a null result when covariance is nonnegative; no zero clipping. The primary Roll (1984), A Simple Implicit Measure of the Effective Bid-Ask Spread supports that definition. The numbers below are synthetic and the arithmetic is author-derived.

Choose the right variant

VariantDefinitionBest useMain limitation
Sample-covariance RollSample lag-one covariance of price changesSmall educational samplesEstimator is unstable in short or trending samples
Population-moment RollPopulation covariance conventionLarge-sample theoryNumerical value differs from sample convention
Extended Roll modelsAdd serially correlated efficient returns or other structureResearch modelsNot equivalent to the simple estimator

Evidence and applicability boundary

This is a research estimator or proxy, not a Rule 605 statistic. Its sample, calendar, currency, and adjustment conventions must be frozen before comparison.

Define the measurement

S^R=2Cov^(ΔPt,ΔPt1)\widehat S_R=2\sqrt{-\widehat{\operatorname{Cov}}(\Delta P_t,\Delta P_{t-1})}

Every price must share one instrument, currency, adjustment basis, and declared quote scope. Intraday measures also need synchronized clocks and correction policy. Daily proxies need one trading calendar and corporate-action basis.

<!-- ENHANCEMENT:F02-TOPIC-SPECIFIC -->

Choose the correct measurement variant

ConventionUse whenWhat changes
Sample price-change covarianceThe canonical educational contractUses paired sample means
Return-based adaptationA study explicitly chooses scale invarianceChanges units and interpretation
Zero-clipped reportA downstream convention requires itMust preserve non-identification separately

Nearest comparison: Effective Spread uses matched trades and quotes; Corwin–Schultz uses daily high–low ranges.

Do not substitute: A null estimate is not a zero spread, and an estimate is not a direct observed quote.

Work the canonical example

Input: {"prices": [100.0, 100.05, 100.01, 100.06, 100.02, 100.07]}.

  1. Successive price changes are 0.05, −0.04, 0.05, −0.04, and 0.05.
  2. Sample lag-one covariance across four pairs is -0.0027.
  3. Roll spread = 2 × sqrt(−covariance) = 0.103923048454, or 10.3886688113 bps at the sample mean price.

See the state before the code

The annotated visual identifies the measurement anchors, active relationship, and failure boundary. Read the labels before comparing the highlighted output.

Roll Spread Estimator annotated measurement

Open the annotated visual at full size · Open the guided playground

Implement the contract

  1. validate ordered positive prices
  2. form successive price changes
  3. pair lagged and current changes
  4. calculate sample covariance
  5. if covariance is nonnegative or within 1e-15 of zero return null
  6. else return 2 × sqrt(−covariance)

The Python and TypeScript references share the same synthetic JSON fixtures and error behavior. They keep raw precision until presentation. Production systems should add fixed-point/decimal prices, feed sequence handling, timestamp normalization, corrections, market-status filters, and licensed source lineage.

<!-- ENHANCEMENT:F02-INTERPRETATION -->

Interpret before aggregating

Read the state before ranking or averaging the value:

CheckQuestion
IdentificationIs lag-one covariance materially negative rather than floating-point noise?
OrderingAre observations chronological and on one adjustment basis?
AssumptionCould fundamental serial dependence dominate bid–ask bounce?

A useful report names the measurement question, variant, scope, clock, unit, weighting rule, valid observation count, and unsupported states. A single headline value is not enough audit evidence.

Validate what the result means

The checks cover canonical arithmetic, a property, a boundary, an invalid state, and cross-language parity. Nonnegative or numerically zero covariance does not identify a real-valued Roll spread and returns null rather than zero. A successful calculation does not prove a profitable strategy, fair execution, or compliance.

StateDecision boundaryRequired response
Negative covariancecovariance < −1e−15Return a real-valued estimate
Near-zero/nonnegative covariancecovariance ≥ −1e−15Return null and not-identified, not zero
Trend or structural breakbounce assumption is implausibleTreat the estimate as model failure, not observed spread

Compare the neighboring methods

Roll needs only prices but relies on stronger stochastic assumptions than quote- or trade-based measures. Quoted, effective, and realized spreads are direct intraday measurements when synchronized quote/trade data exist. Roll and Corwin–Schultz are low-frequency spread estimators with model assumptions. Amihud is a return-to-volume illiquidity proxy rather than a spread.

<!-- ENHANCEMENT:F02-VISUAL-SEQUENCE -->

Use four additional audit views

Roll Spread Estimator calculation map

The map separates eligible evidence, transformation, and reason-bearing output.

Roll Spread Estimator exact decision boundary

The boundary view shows when the method returns a supported value and when it must return an explicit unsupported or bounded state.

Roll Spread Estimator scenario coverage

The matrix confirms that the playground retains five distinct scenarios and 61 deterministic states per scenario rather than relying on a tiny toy example.

Roll Spread Estimator interpretation boundary

The final view separates what the calculation measures from causal, performance, identity, best-execution, and compliance claims it cannot prove.

Evidence boundary

A historical case is deferred until point-in-time identity, licensed source, conditions/corrections, session policy, corporate-action basis, and redistribution permission are archived. NYSE describes Daily TAQ as containing trades, quotes, NBBO, and administrative messages, but its data remain licensed.

Summary and next topic

You can now estimate the Roll spread from lag-one covariance of successive price changes, audit its assumptions, and distinguish it from nearby liquidity measures. Continue with Amihud Illiquidity Ratio.

Primary references

Roll Spread Estimator validation flow

Rendering system map…
ReferencesPrimary sources and evidence notes

Expand the source trail, evidence role, and limitations behind the engineering choices.

Source roles are narrow. Primary sources support definitions and data requirements; all numerical fixtures in this package are visibly synthetic and author-derived.

SRC-01 — Roll (1984), A Simple Implicit Measure of the Effective Bid-Ask Spread

  • Organization or authors: The Journal of Finance
  • Source type: Official rule or original peer-reviewed paper
  • Publication or effective date: 1984-09
  • Version: Source page or paper version at access
  • URL or DOI: Roll (1984), A Simple Implicit Measure of the Effective Bid-Ask Spread
  • Accessed: 2026-07-29
  • Jurisdiction: Research model; market-neutral until implemented
  • Supports: original covariance-based implicit spread estimator
  • Limitations: Does not validate the repository's synthetic numbers or make the selected convention universal.

SRC-02 — NYSE Daily TAQ product description

  • Organization or authors: New York Stock Exchange
  • Source type: Official market-data product documentation
  • Publication or effective date: Current product page; accessed 2026-07-29
  • Version: Source page or paper version at access
  • URL or DOI: NYSE Daily TAQ product description
  • Accessed: 2026-07-29
  • Jurisdiction: U.S. equities
  • Supports: availability of consolidated trades, quotes, NBBO, and administrative messages needed for reproducible intraday measurement
  • Limitations: Product access and redistribution are licensed; the page does not supply a free historical fixture.
liquidity-spreads.ts
export type TradeSide = "buy" | "sell";

function numberValue(name: string, value: unknown, positive = false): number {
  if (typeof value !== "number" || !Number.isFinite(value)) {
    throw new Error(`${name} must be a finite number`);
  }
  if (positive && value <= 0) throw new Error(`${name} must be greater than zero`);
  return value;
}

function sideDirection(side: unknown): number {
  if (side === "buy") return 1;
  if (side === "sell") return -1;
  throw new Error("side must be 'buy' or 'sell'");
}

function quoteValues(bid: unknown, ask: unknown): [number, number, number] {
  const bidValue = numberValue("bid", bid, true);
  const askValue = numberValue("ask", ask, true);
  if (askValue < bidValue) throw new Error("ask must be greater than or equal to bid");
  return [bidValue, askValue, (bidValue + askValue) / 2];
}

export function quotedSpread(bid: unknown, ask: unknown) {
  const [bidValue, askValue, midpoint] = quoteValues(bid, ask);
  const spread = askValue - bidValue;
  const relative = spread / midpoint;
  return {
    model: "quoted-spread",
    bid: bidValue,
    ask: askValue,
    midpoint,
    quoted_spread: spread,
    quoted_spread_relative: relative,
    quoted_spread_bps: relative * 10_000,
    state: spread === 0 ? "locked" : "two-sided",
  };
}

export function effectiveSpread(
  bid: unknown,
  ask: unknown,
  tradePrice: unknown,
  side: TradeSide,
) {
  const quote = quotedSpread(bid, ask);
  const trade = numberValue("trade_price", tradePrice, true);
  const direction = sideDirection(side);
  const effective = 2 * direction * (trade - quote.midpoint);
  const relative = effective / quote.midpoint;
  const quoteSidePrice = direction === 1 ? quote.ask : quote.bid;
  const priceImprovementPerShare = direction * (quoteSidePrice - trade);
  const roundTripSpreadReduction = quote.quoted_spread - effective;
  const tolerance = 1e-12;
  let state = "outside-quote";
  if (effective < -tolerance) state = "through-midpoint-improvement";
  else if (roundTripSpreadReduction > tolerance) state = "inside-quote";
  else if (Math.abs(roundTripSpreadReduction) <= tolerance) state = "at-quote";
  return {
    ...quote,
    model: "effective-spread",
    trade_price: trade,
    side,
    direction,
    effective_spread: effective,
    effective_spread_relative: relative,
    effective_spread_bps: relative * 10_000,
    quote_side_price: quoteSidePrice,
    price_improvement_per_share: priceImprovementPerShare,
    round_trip_spread_reduction: roundTripSpreadReduction,
    state,
  };
}

export function realizedSpread(
  bidAtTrade: unknown,
  askAtTrade: unknown,
  tradePrice: unknown,
  side: TradeSide,
  bidAfter: unknown,
  askAfter: unknown,
  horizonSeconds: unknown = 300,
) {
  const initial = effectiveSpread(bidAtTrade, askAtTrade, tradePrice, side);
  const [futureBid, futureAsk, futureMidpoint] = quoteValues(bidAfter, askAfter);
  const horizon = numberValue("horizon_seconds", horizonSeconds, true);
  const realized = 2 * initial.direction * (initial.trade_price - futureMidpoint);
  const priceImpact = 2 * initial.direction * (futureMidpoint - initial.midpoint);
  const relative = realized / initial.midpoint;
  const impactRelative = priceImpact / initial.midpoint;
  return {
    model: "realized-spread",
    side,
    direction: initial.direction,
    trade_price: initial.trade_price,
    midpoint_at_trade: initial.midpoint,
    midpoint_after: futureMidpoint,
    bid_after: futureBid,
    ask_after: futureAsk,
    horizon_seconds: horizon,
    effective_spread: initial.effective_spread,
    realized_spread: realized,
    realized_spread_relative: relative,
    realized_spread_bps: relative * 10_000,
    price_impact: priceImpact,
    price_impact_relative: impactRelative,
    price_impact_bps: impactRelative * 10_000,
    decomposition_error: initial.effective_spread - realized - priceImpact,
    state: realized < 0 ? "negative-realized" : "nonnegative-realized",
  };
}

function finiteSeries(name: string, values: unknown, minimum: number): number[] {
  if (!Array.isArray(values)) throw new Error(`${name} must be an array`);
  const result = values.map((value, index) => numberValue(`${name}[${index}]`, value, true));
  if (result.length < minimum) throw new Error(`${name} must contain at least ${minimum} observations`);
  return result;
}

function mean(values: number[]): number {
  return values.reduce((total, value) => total + value, 0) / values.length;
}

export function rollSpread(prices: unknown) {
  const priceValues = finiteSeries("prices", prices, 4);
  const changes = priceValues.slice(1).map((value, index) => value - priceValues[index]);
  const lagged = changes.slice(0, -1);
  const current = changes.slice(1);
  if (current.length < 2) throw new Error("prices must create at least two covariance pairs");
  const laggedMean = mean(lagged);
  const currentMean = mean(current);
  const covariance = lagged.reduce(
    (total, value, index) => total + (value - laggedMean) * (current[index] - currentMean),
    0,
  ) / (current.length - 1);
  const meanPrice = mean(priceValues);
  const covarianceTolerance = 1e-15;
  const spread = covariance < -covarianceTolerance ? 2 * Math.sqrt(-covariance) : null;
  const relative = spread === null ? null : spread / meanPrice;
  return {
    model: "roll-spread",
    observation_count: priceValues.length,
    change_count: changes.length,
    covariance_pair_count: current.length,
    lag1_price_change_covariance: covariance,
    covariance_tolerance: covarianceTolerance,
    mean_price: meanPrice,
    spread_estimate: spread,
    spread_estimate_relative: relative,
    spread_estimate_bps: relative === null ? null : relative * 10_000,
    state: spread === null ? "not-identified" : "estimated",
  };
}

export function amihudIlliquidity(
  closes: unknown,
  dollarVolumes: unknown,
  scale: unknown = 1_000_000,
) {
  const closeValues = finiteSeries("closes", closes, 2);
  const volumeValues = finiteSeries("dollar_volumes", dollarVolumes, 2);
  if (closeValues.length !== volumeValues.length) {
    throw new Error("closes and dollar_volumes must have equal length");
  }
  const scaleValue = numberValue("scale", scale, true);
  const returns = closeValues.slice(1).map((value, index) => value / closeValues[index] - 1);
  const ratios = returns.map((value, index) => Math.abs(value) / volumeValues[index + 1]);
  const raw = mean(ratios);
  return {
    model: "amihud-illiquidity",
    observation_count: closeValues.length,
    return_count: returns.length,
    returns,
    daily_illiquidity: ratios,
    illiquidity_per_currency_unit: raw,
    scale: scaleValue,
    illiquidity_per_scaled_volume: raw * scaleValue,
    state: "estimated",
  };
}

export function corwinSchultzSpread(
  highDay1: unknown,
  lowDay1: unknown,
  highDay2: unknown,
  lowDay2: unknown,
  clipNegative = true,
) {
  const h1 = numberValue("high_day_1", highDay1, true);
  const l1 = numberValue("low_day_1", lowDay1, true);
  const h2 = numberValue("high_day_2", highDay2, true);
  const l2 = numberValue("low_day_2", lowDay2, true);
  if (h1 < l1 || h2 < l2) throw new Error("each daily high must be greater than or equal to its low");
  const beta = Math.log(h1 / l1) ** 2 + Math.log(h2 / l2) ** 2;
  const gamma = Math.log(Math.max(h1, h2) / Math.min(l1, l2)) ** 2;
  const denominator = 3 - 2 * Math.sqrt(2);
  const alphaRaw = (
    (Math.sqrt(2 * beta) - Math.sqrt(beta)) / denominator
    - Math.sqrt(gamma / denominator)
  );
  const alpha = clipNegative ? Math.max(0, alphaRaw) : alphaRaw;
  const expAlpha = Math.exp(alpha);
  const spread = 2 * (expAlpha - 1) / (1 + expAlpha);
  return {
    model: "corwin-schultz-spread",
    beta,
    gamma,
    alpha_raw: alphaRaw,
    alpha_used: alpha,
    spread_estimate_relative: spread,
    spread_estimate_bps: spread * 10_000,
    clip_negative: clipNegative,
    state: clipNegative && alphaRaw < 0 ? "clipped-to-zero" : "estimated",
  };
}

export function calculate(topicId: string, inputs: Record<string, any>) {
  if (topicId === "D11-F02-A01") return quotedSpread(inputs.bid, inputs.ask);
  if (topicId === "D11-F02-A02") return effectiveSpread(inputs.bid, inputs.ask, inputs.trade_price, inputs.side);
  if (topicId === "D11-F02-A03") return realizedSpread(
    inputs.bid_at_trade,
    inputs.ask_at_trade,
    inputs.trade_price,
    inputs.side,
    inputs.bid_after,
    inputs.ask_after,
    inputs.horizon_seconds ?? 300,
  );
  if (topicId === "D11-F02-A04") return rollSpread(inputs.prices);
  if (topicId === "D11-F02-A05") return amihudIlliquidity(
    inputs.closes,
    inputs.dollar_volumes,
    inputs.scale ?? 1_000_000,
  );
  if (topicId === "D11-F02-A06") return corwinSchultzSpread(
    inputs.high_day_1,
    inputs.low_day_1,
    inputs.high_day_2,
    inputs.low_day_2,
    inputs.clip_negative ?? true,
  );
  throw new Error(`unsupported topic_id: ${topicId}`);
}
Full-height labplaygroundOpen full screen