The Currency Times

Method

This project publishes no results, so the only thing that can be judged is the process. What follows is that process, stated plainly enough to be argued with.

A theory does not arrive here because it is interesting. It arrives because it can be stated in a way that allows it to be wrong, and it stays only for as long as nothing has shown that it is.

The five gates

1

The claim must be falsifiable before anything is measured

Every entry begins as a null hypothesis, written down in advance. Not "does this indicator help" but a specific statement that data can contradict.

If a claim cannot be put in that shape, it does not enter the list. That single rule removes most of what is written about markets.

2

The data must have been knowable at the time

Every observation carries a release timestamp, and the engine refuses any read dated before it. Not a warning. A refusal.

Raw vendor deliveries are held immutable and every derived layer is rebuilt from them, so a correction upstream cannot silently rewrite a result downstream. The full set of faults this gate exists to catch is set out under Data.

3

The test must not see the answer

Ordinary cross-validation assumes observations are independent. Financial series are not, and overlapping windows leak the future into the past invisibly. Purging removes observations that overlap the test period, and an embargo buffers the boundary.

A single split is not a test, because the split point itself becomes a parameter to optimise. Combinatorial purged cross-validation generates many paths instead of one, which is also what makes the next gate computable at all.

4

The result must be discounted for the search that found it

If enough candidates are tried, the best of them looks good even when every one is noise. The expected size of that false result grows with the number of trials.

So the trial count is recorded before the search, never reconstructed afterwards, and the final figure is deflated for trials, skewness, kurtosis and sample length. The discovery threshold is rather than the conventional two.

One consequence is uncomfortable and worth stating. Any result produced before the trial count was tracked cannot be rescued, because the count is unrecoverable. Those results are not published. They are not published anywhere.

5

The outcome is recorded either way

A theory that fails is written up with the same care as one that survives, and it stays on the list in its own field rather than being quietly removed. Eight entries currently sit as struck off, each with the reason.

This is the gate that costs the most and matters the most, because a research record that only contains successes is not a record.

What is never claimed

No performance figures. No forecasts. No signals. Nothing here is an inducement to do anything, and nothing here is investment advice. The project holds no client money and manages nothing on anyone's behalf.

Where a theory is described as being under test, that is the whole of the statement. It does not imply the theory works, that it is being traded, or that anyone should act on it.

The discipline is not in finding things. It is in refusing to believe the things you find.

Where this comes from

The five gates are a controls framework applied to research rather than to a firm. The founder spent twenty-five years building exactly that kind of framework inside regulated institutions, under consent orders, supervisory examinations and central bank testing programmes. The habits are the same: define the control, evidence it, have it tested by someone whose job is to break it, and record the failures where they can be found later.

Four demonstrations

The argument above, run rather than asserted. Each of these computes in your browser from the mathematics itself.

What gate four looks like

Drag the number of trials. Every candidate in this thought experiment is pure noise with no real skill whatsoever. The figure is what the best of them will show anyway.

E ⁣[maxiNSRi](1γ)Z1 ⁣[11N]+γZ1 ⁣[11Ne]\mathbb{E}\!\left[\max_{i\le N} SR_i\right] \approx (1-\gamma)\,Z^{-1}\!\left[1-\tfrac{1}{N}\right] + \gamma\,Z^{-1}\!\left[1-\tfrac{1}{Ne}\right]
45
1.00expected best score, from noise alone, on five years of daily observations
2.76root of two log N, the standard approximation to the multiplier
2.24implied t-statistic, which is the multiplier itself
Publishable by conventionagainst the usual threshold of two

At forty-five trials the best result already clears the conventional bar for significance while containing no information whatsoever. This is why the threshold used here is three rather than two, and why the trial count is recorded before the search rather than after it.

How much of a correlation matrix is noise

Marchenko-Pastur gives the exact band of eigenvalues that pure noise produces. Anything inside that band is not weak evidence of structure. It is no evidence at all.

50
500
04
0.10ratio of series to observations
0.47lower edge of the noise band
1.73upper edge of the noise band
safestate of the estimate

Push the series past the observations and the matrix becomes singular, which means it cannot be inverted at all. A great deal of applied work quietly sits close to that line.

The effect that gets mistaken for an intervention working

Select the extreme cases on one measurement and look at them again. They will have moved toward the average whether or not anything was done to them.

0.50
+1.76mean of the selected top tenth, first measurement
+0.88expected mean of the same group, second measurement
0.88apparent improvement, caused by nothing

At a correlation of zero the entire selected advantage evaporates. At a correlation of one nothing moves. Every real case sits between, which is why any before-and-after study without a control group is measuring this as well as whatever it intended to measure.

What an estimator returns when there is nothing there

Four hundred series of pure noise are generated in your browser, and the classical Hurst estimator is run on each. There is no memory in any of them. The truth is exactly one half.

1000
truth, 0.50
0.53average estimate on pure noise
0.47 to 0.59where ninety per cent of the estimates land
inconclusivewhat a single reading of 0.58 would mean

The average is only slightly wrong. The spread is the problem. A reading of 0.58 on real data sits inside the range that pure noise produces, so on its own it is not evidence of memory at all. This is why the entry under Long memory reports an interval rather than a number.

NextDatathe inputs, and what is wrong with them