LZLZL/Prediction markets/Research method
FREEBUILD IT F · MethodFlagship

Null models
a profit that cannot beat one is not an edge

2026-08-21 · the cheapest control experiment available

Your strategy made money. But if ignoring the strategy entirely and following a fixed rule made the same money, what you earned is not the strategy's — it is the market's. A null model is that control. It costs almost nothing and eliminates a large class of apparently successful strategies in one pass.

1Profit is not edge

Suppose the market trended one way over some period. Then within it:

ApproachResult
Your strategy+2.5pp
No judgement at all, one fixed side+2.5pp

Identical. Your strategy contributed zero. It looks profitable while actually having stood on the favourable side by accident — and that favourable wind will stop, usually shortly after you size up.

edge = your strategy − the null not "your strategy − 0"

2What to use

A good null has three properties: uses none of your strategy's information, is simple enough to be uncontroversial, and runs on the same sample.

TypeTests
Fixed directionwhether you are earning a one-way market move
Randomwhether you can beat noise at all
Follow the previous outcomewhether you are earning simple serial inertia
Mechanical selection on a public quantilewhether you are earning a structural bias anyone can see
Always take the current pricewhether you beat "no judgement at all, just trust the price"

How to read this tableIt is a generic menu, not my list. The three referred to below do not map onto these five rows — this site does not publish what they were.

The line I killed used three nulls, with the criterion frozen as "must beat all three" — losing to any one meant immediate death, under the verdict "the market wearing a costume".

Basisthis site does not publish what those three nulls were, because their specific form is tied to the strategy. The selection principle transfers though: choose rules that require no judgement to execute.

3★ A null model once saved me

The commoner case is not a strategy losing to a null. It is a null exposing that the whole ruler is broken.

I rescored a full history against real market settlement. The strategy line did not move dramatically. The nulls did:

Old rulerCorrect ruler
A model-free baseline+2.72pp−1.23pp
Another baseline−3.82pp+0.13pp
⚠ A null changing sign means the ruler is broken

A baseline containing no model at all should reflect only the market itself. If changing the basis flips it from clearly positive to clearly negative, the fault cannot be in the strategy — it is in the measurement.

This is nulls' most underrated use: they are both the strategy's control and the instrument's calibration. A strategy moving is attributable to the strategy; a null moving can only be attributed to the ruler.

Full account in the broken ruler.

4⚠ And the reverse: nulls near zero is what "repaired" looks like

After fixing the ruler I retrained against the corrected target. One of the checks that the repair had worked was: contemporaneous nulls should all return to approximately zero.

They did (the contemporaneous nulls all landing between −1.27 and +0.18), while the strategy itself returned four negative cells out of four. Together those make a clean kill: it is not that a broken instrument produced ugly readings — the strategy is dead on the repaired instrument too.

5⚠ Beware "beat the null" being read as success

The opposite error is common. My final read:

CriterionResult
Beat three nullsPASS (all three)
Halves agree in signPASS
Gross excess ≥ 2.25ppFAIL (+1.56pp)
Significance z ≥ 2.0FAIL (1.62)

Beating a null only proves the strategy is not pure noise. It does not prove it covers costs. Between those two sits that line's entire $496.51.

6A related trap: small sample plus fat tail

In one simulation the hourly tenor looked positive (+4.31¢/share). Broken out: one board of six contributed +$667; the other five had a median of −$19.

The mean is meaningless. Ask for the median first, then ask what is left after removing the single best case — two questions as cheap as a null model, and they eliminate just as many false findings.

7One line to keep

Edge is you minus the null, not you minus zero. And nulls do something extra: if the null itself changes sign, what is broken is the ruler, not the strategy. This is the cheapest check in the whole method — it costs almost nothing and can save an entire line.

EvidenceCheck it yourself

Three-null criterion "must beat all three; losing to any one is immediate death", frozen seven days before the read. The specific form of the three is not published.
Nulls changing sign rescoring the full history against market settlement moved one baseline from +2.72 to −1.23pp and another from −3.82 to +0.13pp.
Nulls returning to zero after repair four contemporaneous nulls landed between −1.27 and +0.18pp while the strategy returned four negative cells of four.
Final read beat three nulls PASS, halves agree PASS, magnitude FAIL (+1.56 vs 2.25), significance FAIL (1.62 vs 2.0).
Small sample, fat tail one simulation's hourly tenor: one board of six contributed +$667, the rest a median of −$19.
Checked 2026-08.

NextWhere to go

F · METHOD
The time the nulls changed sign, in full
F · METHOD
Gates and the single read
F · METHOD
Adversarial verification
F · METHOD
Replays must state their bias
This is an educational and research record. It is not investment advice, promises no returns, and offers no personalised trading recommendations. Rules and API behaviour are per the official documentation; this page states when it was checked and both can change without notice. Prediction markets are restricted or unavailable in some jurisdictions — confirm your own before taking part.