algorithmic-trading-anthology-july-2026-placebo-control

Conclusion: Placebo

A deliberately meaningless signal, run alongside the real ones. It found nothing, which is the result it was included to produce — had it scored, the whole screen would be suspect.

Every test, against the significance threshold

persistence-robust p · left is stronger evidence

Filled markers fall below the 0.05 threshold before multiple-testing correction; hollow markers do not. Correction is applied across the whole batch, so an individual marker clearing the line is not a result on its own.

Tests

ticker × horizon

p < 0.05

uncorrected

Best q

—

BH-FDR

Survived

q ≤ 0.10

Method

Hypothesis

Assets with no plausible ENSO channel. If these fire as often as the real hypotheses, the exercise is measuring noise and the framework is mis-calibrated.

Rule

signal = soi_3m # 3-month mean SOI

prior = +0 # none: control

function test(instrument, horizon h):

y = sum(returns[instrument], t+1 .. t+h) # strictly forward

x = signal[t] # known at t

beta, t_stat = OLS(y ~ x) with Newey-West HAC, lag = h + 1

# null: slide x against y by every admissible offset, so

# persistence appears identically under the null

null = [slope(y, roll(x, k)) for k in all offsets]

p = (count(abs(null) >= abs(beta)) + 1) / (len(null) + 1)

return beta, t_stat, p

reject if Benjamini-Hochberg q <= 0.10 across the whole study

Analysis

Calibration

A negative finding is only informative if the instrument can distinguish signal from noise. This set was chosen for having no plausible ENSO channel whatsoever.

It fires at essentially the same rate as the physically motivated hypotheses. The single strongest p-value in the entire first study — 0.0038 — belongs to long-dated US Treasuries, an asset with no agricultural exposure of any kind.

The strongest signal in a study of El Niño and agriculture belonged to the Treasury bond.

Evaluation

Table 1 · Every test in this hypothesis

forward return regressed on the SOI signal

Instrument	Horizon	n	β	HAC t	Shift p	Sign as predicted
^GSPC	1m	676	-0.0002	-0.96	0.298	—
^GSPC	3m	674	-0.0003	-0.59	0.541	—
^GSPC	6m	671	-0.0005	-0.50	0.598	—
^GSPC	12m	665	-0.0014	-0.95	0.403	—
TLT	1m	286	-0.0007	-2.45	0.004	—
TLT	3m	284	-0.0019	-2.67	0.004	—
TLT	6m	281	-0.0024	-1.57	0.054	—
TLT	12m	275	-0.0011	-0.38	0.710	—
GC=F	1m	309	+0.0001	+0.39	0.732	—
GC=F	3m	307	-0.0003	-0.44	0.744	—
GC=F	6m	304	-0.0012	-1.01	0.465	—
GC=F	12m	298	-0.0024	-1.25	0.333	—
MSFT	1m	482	-0.0007	-1.59	0.165	—
MSFT	3m	480	-0.0011	-1.05	0.459	—
MSFT	6m	477	-0.0010	-0.59	0.690	—
MSFT	12m	471	-0.0029	-0.84	0.570	—
^N225	1m	676	+0.0002	+0.96	0.407	—
^N225	3m	674	+0.0007	+1.17	0.325	—
^N225	6m	671	+0.0010	+0.88	0.504	—
^N225	12m	665	-0.0000	-0.02	0.986	—
SI=F	1m	309	+0.0006	+0.96	0.300	—
SI=F	3m	307	+0.0008	+0.56	0.614	—
SI=F	6m	304	-0.0006	-0.23	0.840	—
SI=F	12m	298	-0.0068	-2.24	0.110	—

+ 20 more rows− fewer rows

Shift p is the persistence-robust p-value and is the one that decides. HAC t is shown to demonstrate how far conventional inference overstates the evidence.

Reference

Instruments

last price and one-month change, live from Yahoo Finance

Symbol	Notes
GC=F	Gold — continuous futures on COMEX, in USD.
MSFT	Microsoft Corporation — listed on NasdaqGS, in USD.
SI=F	Silver — continuous futures on COMEX, in USD.
TLT	iShares 20+ Year Treasury Bond ETF — exchange-traded fund listed on NasdaqGM, in USD.
^GSPC	S&P 500 — market index, in USD.
^N225	Nikkei 225 — market index, in JPY.

Sample

1970–2026 monthly

Sources

Queensland Department of Agriculture and Fisheries. Southern Oscillation Index. The Long Paddock. Monthly, 1876–. www.longpaddock.qld.gov.au/soi/soi-data-files

Yahoo Finance. Historical market prices. Adjusted close, month-end. Retrieved with yfinance. finance.yahoo.com

Revision history

Rev	Stage	Status	Change
01	Draft	—	Direction and criteria fixed before the experiment ran
02	Tested	Placebo	24 tests; 2 below p=0.05, none surviving FDR