Abstention is a decision
The first thing a good options trader learns is when to sit out. The second thing they learn is that "sitting out" is itself a position — the position of being in cash, accruing zero, vs the position of being in the worst trade of the day. Most "AI" tools are bad at sitting out because they're scored on what they say, not on what they refuse to say.
Why abstention is hard
If your training loss rewards predicting the right side of a move, your model learns to always predict some side. NO_TRADE looks like cowardice to the loss function. The model finds an answer even when there isn't one.
This is the failure mode behind every "AI says BUY" alert that goes off in chop, in earnings windows, in macro panic. The model wasn't designed to abstain — it was designed to argmax.
How we score abstention
We have a metric we call the brutal pass rate. Across a hand-curated set of "should-have-said-NO_TRADE" setups, what fraction does the model correctly abstain on?
Examples in the brutal set:
- The 30 minutes around a CPI print on a high-vol day.
- SPY/QQQ disagreement > 0.6% (sentiment-NEUTRAL by our classifier).
- A ticker 2 days from earnings with elevated IV.
- A pattern setup with rel-volume below 0.5 (failed breakout signature).
A model that scores 80% win rate but 20% brutal pass rate is a model that will blow up the account on bad days. A model that scores 60% win rate but 90% brutal pass rate is the one we ship.
The veto specialist
We train a dedicated veto head whose only job is to fire NO_TRADE on a parallel dispatch — separate from the chart-direction head's argmax. The cluster combines both. The veto head doesn't care about being right on direction; it cares about being right on "is this setup actually scoreable at all?"
The result: the cluster refuses far more setups than the chart head alone. Some of those refusals would have been winners. We're fine with that. Smaller variance, lower max drawdown, longer survival.
An AI that always gives you an answer is selling you certainty it doesn't have.