Most of what passes for market signal is either tautology, narrative, or overfit. A durable signal has to survive three tests: it has to be mechanistically grounded, falsifiable in advance, and decision-relevant in the size that matters to the decision-maker. Most published indicators fail at least one of these.
What is not a signal
Several common artifacts get treated as signals and shouldn’t be:
- Story-shaped correlations. Two series moved together for a few years. Without a mechanism, this is an observation, not a signal.
- Restated truisms. “Liquidity matters.” “Positioning matters.” “Sentiment matters.” These are framings, not signals. They become signals only when they specify which liquidity, what positioning, and which sentiment, measured how.
- Indicators built on retrospective rebalancing. Anything tuned on the same window in which it is evaluated is a description, not a prediction.
- Lagging confirmations. A signal that fires after the move is not a signal. It is a chronicle.
What a useful signal looks like
A signal earns its name when it has four properties:
- A mechanism. A reason, however imperfect, that connects the observation to a behavioural or institutional driver. Mechanisms can be wrong, but they can be debated.
- A falsification rule, specified in advance. “If X reaches level Y by time Z and the expected effect does not materialise, the signal is wrong.” Without this, no signal can decay; every miss gets rationalised.
- Sufficient lead time. It has to fire early enough to be actionable in the relevant size. A signal that fires at the inflection point is indistinguishable from price itself.
- Robustness to small specification changes. If the same idea, measured slightly differently, produces materially different histories, the signal is fit to its own definition, not to a phenomenon.
Reflexivity is the hardest case
The signals that matter most often have a reflexive component: the act of observing or trading on them changes the underlying. Positioning data, flow-based indicators, and narrative momentum measures all share this property. They can still be useful, but they require a structural adjustment: the question is not “what does this signal say?” but “what does this signal say given that many participants are now watching it?”
A useful test: if everyone with access to the signal traded on it simultaneously, would the original mechanism still hold? If not, the signal is borrowed from its own popularity, and its half-life is shorter than the backtest suggests.
Decision relevance
A signal that is mechanistically clean, statistically robust, and well falsified can still be useless if it is irrelevant at the size and horizon of the decision. The discipline is to specify the decision first and then ask whether any candidate signal is informative at that horizon and that level of conviction. Working in the other direction — collecting signals and looking for decisions — produces analysis that is interesting but rarely usable.
What we use this for
Internally, we apply this filter as a routine check on any indicator before it enters research output:
- What is the mechanism?
- How would we know it has broken?
- Is it leading, coincident, or lagging at the decision horizon?
- Has it survived the same specification across different windows?
- Are we watching it because it is informative, or because others are watching it?
The output of that filter is a much smaller set of signals than we started with — which is, on its own, useful information about how much of the public commentary deserves the weight it is given.