Is market conductance a forecastable state variable? A free-data falsification study of the impact operator in gold and nine other markets

codedpro; ITMaster; Code Nest

Free market data sources used in this study

Every endpoint, what it costs, what it covers, and which sources that were free are now blocked. No third-party data is redistributed — collectors fetch at run time.

Sources that work

All free, no API key required except where noted. Probed 6 September 2026.
Source	Role	Coverage	Status
CFTC Disaggregated Commitments of Traders — Socrata API at `publicreporting.cftc.gov`, gold contract `088691`	Holder base	1,056 weeks, 2006–2026	OK
Daily OHLCV for ten ETFs — GLD, SPY, QQQ, TLT, HYG, SLV, USO, FXE, EEM, GDX	Instruments	4,640 days each, 2008–2026	OK
LBMA gold price — official PM fix feed	Price	14,676 days, 1968–2026	OK
CBOE GVZ — gold implied volatility index	Control	4,594 days, 2008–2026	OK
Federal Reserve FOMC calendar — meeting blocks and statement links	Event dates	148 scheduled decisions, 2008–2026	OK
Dollar index, 10-year Treasury yield	Controls	2001–2026	OK

Sources that are no longer freely accessible

Two corrections to the standard data inventory, both discovered by probing rather than assumed:

Source	Role	Status
COMEX warehouse stocks — registered versus eligible, `cmegroup.com/delivery_reports/Gold_Stocks.xls`	Absorption	403 — bot-blocked
LBMA vault holdings — seven custodians plus Bank of England	Absorption	404 — URLs moved
World Gold Council Goldhub	Absorption	registration
FRED CSV endpoint	Control	rate-limited

Together the first two are the entire absorption stack — the denominator of Λ, including the registered-versus-eligible split that makes COMEX unusual. That half of the framework could only be tested through weaker proxies (open interest and trader count), which is a material limitation on the result.

A warning: Yahoo GC=F historical volume is broken

Not merely noisy — unusable. Median daily volume by year runs 56 to 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 fine; volume is not. Any study that computes turnover, Amihud illiquidity or impact from this series will produce nonsense. This study measures on GLD, whose consolidated NYSE Arca volume is clean across nineteen years: median 6–15 million shares, no zeros.

Extracting FOMC decision dates correctly

Each FOMC policy decision publishes exactly one statement, so the Federal Reserve's own statement links are an authoritative, non-derived list of decision dates. Two traps:

Exclude unscheduled meetings. Emergency actions happen because markets are stressed. The Fed labels them — "(unscheduled)", "(cancelled)", "Conference Call", "notation vote" — so exclusion follows their taxonomy, not a judgement call.
Watch month-straddling meetings. A two-day meeting spanning a month end is headed "April/May 30-1" or "Jan/Feb 31-1", so a naive month filter silently drops it.
Not every monetary<date> link is a decision. Framework releases use the same URL pattern. Cross-checking against press-conference links and the Wednesday decision convention removes them.

Reproducing the study

git clone https://github.com/codedpro/market-conductance
cd market-conductance
make setup      # venv and pinned dependencies
make all        # collect, measure, engine, cot, events, generalize
make test       # calibration: size, power, point-in-time discipline

Roughly fifteen minutes, total cost $0. Random seeds are fixed. Collectors write immutable raw snapshots with provenance attached, and every series carries both its reference and release dates. Machine-readable results land in out/ as JSON.