How tauTrader signals are produced and validated

For one asset and one movement threshold, the model estimates whether the asset's price will reach at least that threshold above its reference close, and hold it for two consecutive five-minute closes, at some point within the forecast window. A signal is issued when the model's output reaches that model's frozen decision threshold.

A signal is not a trade instruction, a price target, a return, or a position.

tauTrader currently covers 6 US equities:

Each asset has 6 movement thresholds — 0.5%, 0.7%, 1.0%, 1.25%, 1.5%, 1.75% — and each of those is a separately trained and separately validated model. They evaluate independently, so an asset can signal at several thresholds on the same day. When that happens you will see the whole ladder that fired.

A higher threshold is a larger move, not a stronger signal. The thresholds are not ranked against one another, and one firing does not make another more likely.

The window is 78 regular-session five-minute bars measured from the reference close, which is through the close of the next regular trading session. It spans the overnight gap, so overnight moves count toward the threshold.

Because the window is defined in trading sessions rather than clock hours, it is longer than a calendar day across weekends and market holidays. A signal stops being current when its window closes.

All models forecast upward movement. There is no downward model and no downward signal.

The absence of a signal is not a bearish forecast. It carries no directional claim at all — only that no model reached its threshold for that asset today.

Days with no signals at all are normal and expected, not an outage.

One frozen family of models was evaluated once, on data it had never seen. Decision thresholds were fixed before that evaluation and were not touched afterwards; no model was retrained, retuned or reselected on the basis of the result.

Precision is the share of signalled bars on which the move occurred. Base rate is the share of all evaluated bars on which it occurred, signal or not. Lift is the difference between them, in percentage points — the part precision alone does not tell you. Precision falls as the threshold rises, because a larger move is a harder question; the base rate falls with it, which is why lift is the comparable quantity.

The model's numeric output is a decision score, not a calibrated probability that the move will occur. We publish the signal, not the score.

There is no confidence tier, and no signal is presented as more or less confident than another. Past validation performance does not guarantee future results.

The product runs one frozen family of models at a time, identified by a short version stamp (currently 0edfd7f2). Within a version, models are never swapped, retuned or reselected, and decision thresholds never move.

A new family is a deliberate, versioned event, published with its own held-out evaluation. Signals issued under a previous version keep the version that produced them.

Model family identity 0edfd7f2fa4a5d7d08993040ac4740430ca27de9b724180759feaa21bae0e8da

Signals are not resolved against what actually happened. tauTrader does not currently publish per-signal outcomes, realised returns, or a running hit rate — building that requires an outcome-tracking pipeline the product does not have yet, and we would rather say so than show an empty statistic.

tauTrader is experimental and informational. It is not financial advice, and markets involve risk.