Why Can Option Execution Costs Be Lower than Midpoint Estimates?¶

Bottom line: A quote midpoint may be a poor fair-value benchmark when traders choose favorable execution times. Correcting that benchmark can reduce estimated effective spreads; it does not justify discounting backtest fill prices, slippage or commissions.

Source and version¶

Dmitriy Muravyev and Neil D. Pearson (2020), Options Trading Costs Are Lower than You Think, The Review of Financial Studies 33(11), 4973–5014.
DOI · Publisher record
First published online: February 10, 2020; issue: November 2020. Filed under 2020 Q1, not the issue quarter or the 2026 reading date.
Read version: the complete 42-page published article, including its title, abstract, main text, Tables 1–10, printed Appendix A.1–A.3 and references. The separate Internet Appendix was not obtained.
peer_reviewed describes publication; read describes source review, not independent replication.

Question and data¶

How can options trade actively despite wide quoted spreads? The authors argue that executions need not arrive randomly: traders may buy when fair value approaches the ask or sell when it approaches the bid. Measuring these trades against the current midpoint can then overstate costs.

Section 1, pp.4978–81, uses Nanex OPRA trades and one-minute quotes, SIP stock quotes and CRSP for options on S&P 500 constituents during January 2004–December 2015. The filtered sample contains 620,916,282 trades and 1,242,893 stock-days. Quotes are snapshots, not every order-book update, and the data retain only contracts traded that day. Low-priced trades, uncertain trade directions, unavailable or extreme predictions, and the first and last five trading minutes are excluded.

These are equity options on constituent stocks—not SPX index options or a contemporary 0DTE sample.

Method: change the benchmark, not the fill¶

Let \(p_i\) be the execution price, \(m_i\) the pretrade midpoint, and \(B_i=+1/-1\) the buy/sell sign. Conventional and adjusted relative effective half-spreads are

\[ ES_i=\frac{(p_i-m_i)B_i}{m_i}, \qquad AdjES_i=\frac{(p_i-\widehat V_i)B_i}{m_i}. \]

Their difference is the estimated timing benefit:

\[ ET_i=\frac{(\widehat V_i-m_i)B_i}{m_i}. \]

The fair-value proxy is the forecast midpoint ten minutes ahead. Predictors include a BSM price computed from the previous thirty minutes of IV and the current stock price, lagged stock/option changes, and exchange participation at the NBBO (Sections 2.1–2.2, pp.4983–89). This is an estimated benchmark, not directly observed true value.

Two exercises must remain distinct:

Main cost estimates use quote regressions fitted separately for each contract and trading day. Their coefficients are not automatically available at every execution time that day.
Table 4 separately tests next-month forecasts using previous-month grouped coefficients. Mean OOS \(R^2\) is 5.5% / 7.0% for Models 1/2. That tests predictability, not independently validated OOS execution costs or strategy profits.

Author-reported costs¶

Table 5, p.4997, reports stock-day averages. Percentages use the pretrade option midpoint as denominator, not the underlying price, margin or account capital.

Half-spread measure	% of midpoint	Cents per share
Quoted	8.6%	12.9
Conventional effective	6.8%	9.7
Adjusted, across all traders	5.0%	6.8
Algo, inferred execution timers	2.5%	2.8

Overall adjusted costs are about one-quarter below the conventional estimate. The algo estimate is lower still, but it describes a latent group, not identified account-level expenses. Under zero conditional mean and median timing for nontimers, positive timing for timers, and the stated group-cost assumption, the authors estimate

\[ q=1-2\Pr(ET<0), \qquad AlgoES=\frac{AdjES-(1-q)ES}{q}. \]

The average inferred timing share is 38.2% (Section 3.1, pp.4994–96; Appendix, pp.5010–12). This does not identify 38.2% of accounts as bots; the mean, median and group-cost assumptions matter.

Table 8, p.5002, reduces conventional ten-minute price impact from 1.526% to 0.644% with Model 2. This is a model-adjusted statistic after removing predictable movement, not randomized causal identification of an order's impact.

Why this is not a plug-in backtest discount¶

Section 5, pp.5004–08, illustrates cost sensitivity with monthly short near-ATM straddles held to expiry. It does not simulate our order waiting, cancellations, queues, capacity or joint-leg fills.

Entry costs are imputed from executed trades in similar-delta, similar-maturity options that day, not necessarily trades in the selected straddle.
Trade direction, timing and trader type are inferred. As a reader-inferred caveat, fair-value errors correlated with trade direction can affect interpretation.
The illustration deducts entry spreads, not all price impact, commissions, funding, assignment or hedging costs. Returns use initial straddle premium, not margin or maximum loss.
Table 10 panel A's return deductions do not reconcile item by item with its displayed cost statistics, and some auxiliary high–low entries differ from the displayed endpoints. This reporting ambiguity remains unresolved; we do not turn it into a calibrated cost schedule or a replication claim.
Same-sample predictor selection, same-day fitted costs and the historical setting do not become independent OOS strategy evidence merely because Table 4 tests price forecasts.

Research implications¶

These are proposals to test, not production conclusions:

Keep fair-value-relative execution quality separate from cash PnL at actual fills. A different benchmark does not increase premium received.
Test timing using only point-in-time coefficients, IV, quotes and signals. Compare against immediate execution including waiting risk, missed trades and exposure.
Validate costs at the package level. Historical single-leg average half-spreads are neither executable combo quotes nor guaranteed simultaneous fills.

Research boundary

This is AI-assisted original source analysis, not data/code replication. The restricted PDF, extracted text and figures remain local and are not distributed with this public page.