Teaching a model to read the tape
Research into how a model can learn what a trader notices and surface those patterns across markets, while the people keep the judgment.
Narrated in the writer’s own voice · 3:06
Contents4 sections
Start with something familiar. The relative strength index is a simple, well-understood read of momentum: cheap to compute, easy to reason about, and an honest baseline for a lot of tape-reading work. It is also a useful on-ramp for the idea we care about, because an indicator is really just a compressed opinion about the tape. Someone decided that recent gains relative to recent losses say something worth watching, wrote that intuition down as a formula, and now a line on a chart stands in for a judgment. RSI is legible precisely because that judgment is fixed and small.
Labeling attention, not outcomes
The move we are exploring is to stop hand-tuning one indicator and instead label the moments that actually mattered — the setups we leaned into, the ones we passed on, the ones we wish we had caught. Then we train a model to recognize the context around those moments rather than the outcome that followed. The features are a blend of the reads a trader already carries in their head: momentum indicators like RSI and its relatives, volatility, order flow, time of day.
The target is not a price forecast. It is the trader's attention.
We are asking the model to learn what we notice, not what the market is about to do.
What it is not
Be clear about what that is and is not. It is not a crystal ball, and we would not trust one if it claimed to be. A model that learns a person's reads also inherits that person's blind spots — the setups they systematically ignore become setups the model ignores too. So it stays a signal to a person, never an autonomous decision. It earns trust the same way anything else in the stack does: proven out-of-sample, costed honestly against real fills and fees, and retired the moment its edge decays rather than nursed along out of attachment.
Leverage on attention
The reason to bother is narrow and practical. An experienced trader can only watch a few names closely at once; attention is the binding constraint, not insight. A model that has learned those reads can watch the whole board and raise a hand — surfacing the setups we would have flagged ourselves if we had been looking in that direction at that moment. It does not decide. It points. That is leverage on attention, and attention is the thing that runs out first.
Where it stands
None of this is a model that trades. The goal is smaller and, we think, more durable: a model that notices the way we do, so two people can be in more places at once. It remains research, honest about its limits, and it lives or dies by the same process discipline as everything else we run. If it stops helping us make better calls, it comes out.