When the Classifier
Doesn’t Know —
That’s the Signal.

A neural network classifier trained on $GLD not only detects crowded trades — its zone of maximum uncertainty anticipates regime inflections that rolling volatility confirms too late.

What Most Positioning Models Fail to Capture

Standard positioning frameworks — historical percentiles, flow z-scores, net COT — all answer the same question: where is positioning right now? They are descriptive. They capture the level but not the market structure that produces it.

There is a different, more valuable question: how well does the current environment fit the expectations the model built from history? When returns start behaving in ways the classifier hasn’t seen before, that divergence is information — and it arrives before the trade reverses.

This is exactly what a neural network classifier trained on implicit positioning captures. Its zone of maximum uncertainty — where it assigns nearly equal probabilities to all three regimes — is not a model failure. It is the cleanest signal the model produces.

“Markets don’t reverse on news. They reverse on positioning. And positioning leaves a trail in price — if you know how to read it.”

The Positioning Proxy: Built From Price

Without access to real-time COT data, we infer positioning from price structure. The logic is precise: if $GLD has risen consistently over 63 trading days (~3 months), there is a mass of accumulated buyers sitting on unrealized gains. That is a structurally crowded trade. The reversal is not caused by the catalyst — it is caused by the structure.

The Architecture: MLPClassifier Over the Positioning Space

The classifier follows the same architecture Matthew Dixon demonstrates in his 2019 notebook on deep classifiers in finance — but instead of separating synthetic half-moons, it separates real market regimes in the z-score × realized volatility space.

The advantage over a binary logit is structural: the decision boundary in this feature space is non-linear. Market regimes are not separable with a straight line. The MLP with ReLU activations learns that curvature directly from data — no analyst encodes it.

The Transformation: Input Space → Hidden Layer

Dixon’s 2019 notebook demonstrates that a MLP’s hidden layers don’t just classify — they transform the feature space to make linearly separable what wasn’t in the original input. The same happens here with real $GLD data.

With a final hidden layer of 2 units (directly plottable), we can visualize exactly what the network did: it takes a space where the three regimes overlap, and builds a representation where the clusters are linearly separable.

Figure 2. Transformation from input space (left) to the 2-unit hidden layer space (right). In the input, the three regimes overlap — no linear classifier separates them. After the ReLU transformation, clusters reorganize into linearly separable zones. Replicates Dixon’s (2019) make_moons analysis, applied to a decade of real $GLD data.

.

The Signal Few Exploit: The Maximum Uncertainty Zone

Most classifier users only read the final output: “regime = Crowded Long”. They ignore the internal class probabilities. That is a mistake with real operational cost.

When the classifier assigns probabilities close to 33%–33%–33% across the three regimes, it is not confused — it is detecting that the market is at a structural transition point. That state of maximum model uncertainty consistently precedes regime changes.

Figure 3. Evolution of $GLD price over time — animation built with an expanding window from 2010. Each frame advances 40 trading days. The accumulation and reversal pattern is visible in price structure, not price level. Data: Yahoo Finance · gif generated with imageio.

Why This Complements (and Doesn’t Replace) the Percentile