Questions about market conductance and price impact
Direct answers on the impact operator, Kyle's lambda, positioning data, free data sources and reproduction.
What is market conductance?
Market conductance is the operator that converts a shock into price movement, written G in the relation ΔP = G × ε, where ε is the shock. It is equivalent to Kyle's lambda — the price impact per unit of order flow. Forecasting conductance means forecasting how hard a market will move for a given piece of news, rather than forecasting the news itself.
Is market impact a constant or a state variable?
It is a state variable. Measured conductance is strongly persistent: Ljung-Box(22) = 1021.5 for gold against an iid-bootstrap null 99th percentile of 40.0, variance ratios rising from 1.34 at one week to 6.82 at one quarter, and a HAR out-of-sample R² of +7.3%. The constant-impact model is rejected at p < 0.0001 in all ten markets tested.
Does the conductance result apply only to gold?
No. The identical signature appears in all ten liquid markets tested — gold, S&P 500, Nasdaq, long Treasuries, high-yield credit, silver, crude oil, the euro, emerging-market equity and gold miners — each at p < 0.0001. High-yield credit shows the strongest out-of-sample forecastability at +10.1%; gold sits mid-pack. Because the effect is universal, a mechanism specific to gold's holder base cannot explain it.
Does Commitments of Traders positioning predict volatility or market impact?
No. Managed Money positioning from the CFTC Commitments of Traders report — the standard proxy for margin-sensitive, forced capital — does not forecast conductance at any horizon, with every t-statistic below 1.2 in absolute value. The only Commitments of Traders variables with predictive power are generic market depth (open interest, t = −2.88) and breadth (trader count, t = −3.18), both with the negative sign that absorption capacity implies.
Does conductance predict how markets respond to scheduled events like FOMC?
No. On 137 scheduled FOMC decisions between 2008 and 2026, pre-event conductance carried a coefficient of −0.123 (t = −0.59) on the size of the decision-day move, adding nothing beyond option-implied volatility. The theory predicted a coefficient near +1.0, which lies 5.4 standard errors outside the 95% confidence interval of [−0.53, +0.29]. Implied volatility priced the same risk almost exactly right, with a coefficient of 1.005.
What is the market impact exponent, and why estimate it rather than assume it?
The exponent α governs how price impact scales with volume in log|r| = log G + α × log v. The square-root law of market impact assumes α = 0.5; Amihud illiquidity assumes α = 1.0. Estimating it gives 1.032 for gold and a range from 0.41 (euro) to 1.36 (S&P 500), so no single assumed exponent would have been correct. Using the empirical α also makes log conductance orthogonal to log volume by construction, so persistence cannot be an artefact of volume.
How do you tell a real impact operator from ordinary volatility clustering?
Use scheduled events. If persistence in measured impact were just clustered news, then on a date fixed years in advance — where arrival carries no information and the surprise size is near-unpredictable — pre-event conductance should add nothing beyond what options already price. If it were a genuine transmission operator, it should forecast the response with a coefficient near 1. In this study it added nothing, and pre-event conductance lost significance precisely on scheduled days (t = +1.57) while loading strongly on ordinary days (t = +7.98). That ordering favours volatility clustering.
What is a kill condition?
A falsification criterion written down before any data is collected, specifying in advance what result would end the project. Four were pre-registered here. Pre-registration is what stops the decision being made in the moment by whoever most wants the theory to be true.
Is Kyle's lambda the same as Amihud illiquidity?
They are closely related measures of price impact. Kyle's lambda is the structural coefficient in r = λ × flow. Amihud's ILLIQ is a daily proxy, absolute return divided by dollar volume, which corresponds to assuming an impact exponent of 1.0. This study nests both by estimating the exponent rather than fixing it.
Can this be traded?
No trading strategy is supported by these results. The signed-asymmetry test, which was the route to a directional edge, produced no significant drift at any horizon. Adding the holder-base measure to a volatility forecast reduced out-of-sample accuracy in five of six specifications. This is empirical research and contains no investment advice.
Is COMEX warehouse stock data still freely available?
Not by direct download. The COMEX gold warehouse file at cmegroup.com/delivery_reports/Gold_Stocks.xls now returns HTTP 403 behind bot protection, and LBMA's vault-holdings URLs return 404. Together these are the entire absorption side of the framework, which is why that half could only be tested through weaker proxies.
Is Yahoo Finance GC=F volume data reliable?
No, not historically. Median daily volume by year runs 56–905 contracts against a maximum near 200,000, with 41 zero-volume days — the series alternates between the front-month aggregate and a near-dead contract. Price is usable; volume is not. This study measures on GLD, whose consolidated NYSE Arca volume is clean across nineteen years.
How is point-in-time discipline enforced?
Commitments of Traders data is a Tuesday snapshot published the following Friday at 15:30 ET. Every series stores both a reference date and a release date, and all joins use release date only. An automated test asserts that no row is ever used before the day it was published. Timestamping by reference date instead produces a backtest that looks excellent and means nothing.
How was the null distribution constructed?
Under the joint hypothesis that conductance is constant and that volume co-moves with volatility only through information arrival, log conductance is iid. The null distribution comes from 2,000 iid bootstrap resamples of the empirical series, preserving its exact fat-tailed marginal while destroying time structure. The identical statistic pipeline runs on real and simulated data, so anything the pipeline itself induces appears in the null too.
Was the test powerful enough to detect the predicted effect?
Yes. The HAC standard error on the pre-event conductance coefficient is 0.208, so the theory's own prediction of +1.0 would have appeared at roughly five sigma had it been present. This is an informative null rather than an underpowered one, and it is stable across pre-event windows ending 2, 6 and 11 days before the event.
Can I reproduce these results?
Yes, at zero cost. Clone the repository, run make setup then make all. Every number regenerates from public endpoints in roughly fifteen minutes; no third-party data is redistributed and random seeds are fixed. The code is MIT licensed at https://github.com/codedpro/market-conductance.