Article 4 of 7 — Building Momentum Systems Without Fooling Yourself. Previous: Labelling: Triple Barriers in Volatility Units · Next: Regime Filters as Features, Not Conditions

The labels are defined, registered, and frozen. The natural next move, the one every trader’s hand reaches for, is to build a strategy. I do not do that yet, and this article is about why. The first thing that runs on the data is not a strategy. It is a map: a small family of deliberately ordinary trend-entry primitives, run identically across all twenty-five markets, with a fixed exit structure, no filters and no optimization. The problem the map solves is simple to state and easy to skip: before I decide what to improve, I need to know what unimproved momentum actually looks like on this universe, and I need to know it in a form that does not already contain my preferences.

Six primitives, one exit, no knobs

The family is textbook by design. Two kinds of moving-average crossover, one that reads the direction of an average’s slope, a channel breakout, a rate-of-change signal, and a regression-slope signal. Anyone who has opened Kaufman knows all six. Each is a separate C# primitive in the signal layer, new signals are new primitives, never edits and each does one thing: it proposes a side. It does not decide how long to hold, where to get out, or whether the market is worth trading today. Those are other layers, and for the baseline they are held constant.

The exit structure is the triple barrier from Article 3, unchanged, because the whole point of specifying it in volatility units and registering it a priori was to make it reusable without per-instrument fiddling. A probe fires, the barrier geometry takes over, the trade resolves at the stop, the target or the horizon. Every probe gets the same treatment on every market. Nothing is tuned per instrument; nothing is tuned at all. The parameters are the ones a textbook would suggest, chosen before any run and recorded alongside the labelling specification.

I want to be blunt about how boring this is, because the boredom is the feature. A baseline that has been polished is no longer a baseline; it is a first strategy wearing a baseline’s clothes, and everything downstream inherits its hidden choices. “Baseline before filtered” is a rule in the framework for exactly the reason the “Illusion” posts described. If the reference point has already absorbed three or four judgment calls, a smarter exit here, a market dropped there, then every later comparison is a comparison against a moving target that I moved. The unfiltered baseline is the zero of the ruler. You do not get to carve the zero.

One trial, not one hundred and fifty

Six probes across twenty-five markets, long and short, is a lot of backtests. It would be easy to read that as a lot of trials, and in the sequential-optimization world it would be: each would be a candidate, each candidate a chance to keep the winner. The register from Article 2 treats the batch differently, and the distinction matters.

The batch is registered as a single family-level trial. Its hypothesis is not “probe X works on market Y”. Its hypothesis is descriptive: “textbook trend-following, applied uniformly, has this shape across the universe”. The unit of inference is the family. What is committed before the run is the list of probes, the universe, the exit, and the metric; what the run produces is one object, a table of outcomes, read as a whole.

This is not an accounting trick to keep the trial count low. It is what stops the map from turning into a menu. The moment I look at the table and pick the best probe for each market, I have silently run a selection over a hundred and fifty backtests and reported the winners, the exact mechanism the “Illusion” posts dissected, with the noise lottery drawn twenty-five times. The family-level registration makes that reading illegitimate by construction: the table was never registered as a set of candidates, so nothing in it can be promoted to a strategy without a new, separately registered hypothesis that says what it is testing and why.

So the table is read as a map. Three things are extracted from it, and none of them is a number to be reported as performance. The first is ranking: which probes sit above which others, market by market, and whether the order is the same anywhere. The second is breadth: how many markets a probe is on the right side of, judged against chance — more on that below. The third is sign: whether a probe’s contribution is positive at all, on which side, and whether that survives moving from one instrument to its neighbour. Ranking, breadth, sign. Not “which one to trade”.

What the map says

Two things stood out, and both of them are structural rather than numerical, which is the only kind of finding a baseline should be trusted to give.

The first is a long/short asymmetry on US equity index futures. Every probe in the family, without exception, does its useful work on the long side, and every probe destroys returns on the short side. This is not a discovery about momentum; it is the equity risk premium showing through a trend-following lens. A symmetric system on a market with a persistent drift is fighting the drift half of the time, and the textbook entries are too crude to know it. What makes the finding useful is the control: on the German index the asymmetry is absent. Same probes, same exit, same specification, the structure of the outcome is different, which means the structure belongs to the market and not to the method.

The design implication is direct. For US indices, a symmetric momentum system is the wrong object. The candidates are either long-only, which is honest but crude, or the “gate + direction” structure that Article 5 develops: a regime gate that decides whether the market is tradeable at all, and a probe that decides direction only once the gate is open. What is not on the table is bolting a “no shorts on US indices” condition onto each probe after seeing the result. That would be a filter chosen on the outcome, and I already know how that story ends. The asymmetry enters the framework as a registered observation that shapes which hypotheses are admitted next, not as a patch.

The second finding is that cross-market ranking is unstable. The breakout probe is among the top two on US indices and is the worst of the six on the German index. The others reshuffle less dramatically but they reshuffle. I looked for a story, liquidity, session structure, index composition and I do not have one I would defend. What I have is a rule: a probe ranking obtained on one market says almost nothing about the ranking on another, even a closely related one. So “pick the entry that works best here” is not a decision rule I am allowed to use, and any strategy that depends on that choice is depending on a coin.

The yardstick for breadth

Reading “breadth” off the table needs a reference, and the reference cannot be zero. A probe that is positive on a majority of markets sounds impressive until one asks how many markets a coin-flip entry with the same exit structure would be positive on. Equity indices are correlated with each other; energy contracts move together; a good year for trend on one rates market is usually a good year on three others. Counting markets as if they were independent inflates breadth in exactly the way that flatters the researcher.

Following Davey, the baseline batch is accompanied by a random-entry benchmark: entries drawn at random, resolved by the same triple barrier, run across all twenty-five markets simultaneously so that the real cross-market correlations are preserved in the benchmark’s own distribution. The output is a distribution of breadth under chance, and the probe family’s breadth is read against it. It is a diagnostic, not a gate. A probe that does not separate from the random benchmark is not “rejected”; it is marked as showing nothing that chance would not show, which is a fact about the map, not a verdict. Article 7 treats the benchmark properly, how the distribution is built, what it is and is not evidence for. Here it plays the humbler role of stopping me from reading “profitable on most markets” as if it meant something on its own.

The question the baseline cannot answer

There is one thing the map cannot tell me, and I would rather say so now than have it surface later as a surprise. The baseline is run over roughly twenty years of daily data, and the effects it targets are assumed to live at a horizon that is fixed across that span. I am not sure that assumption holds. There is a reasonable case that the horizon of exploitable trend on liquid futures has shortened as more capital arrived to trade it, and if that is true, a probe that is on the right side over the full sample may be on the right side of an effect that was stronger in the first half than in the second. The baseline aggregates over time, and aggregation hides drift. Stationarity of the horizon is an open question in the framework, registered as such, and nothing in this article should be read as if it were settled.

What the baseline has done is narrower and more useful than a strategy would have been. It has fixed the reference point that every later comparison will be made against; it has produced two structural facts about the universe that constrain which hypotheses are worth admitting; and it has cost exactly one trial. The next article uses those facts. Regime primitives — efficiency, persistence, choppiness, volatility state, enter not as conditions bolted onto these probes, but as features for a model that decides when the gate is open, which is the only version of “filtering” this framework allows.

References

Davey, K. (2014). Building Winning Algorithmic Trading Systems. Wiley. Kaufman, P. J. Trading Systems and Methods. Wiley.

quanthedgeai, the research arm of algosworksai. Systematic research on mid-frequency instruments.