My Z-score model got destroyed by black swans so I built a neural net to fix it

Following up on my last post about mean reversion engines. Going to be completely honest my initial Z-score model got absolutely cooked during the last few high volatility macro drops.

A Z-score of 3.5 in a normal choppy market is a beautiful fade. The reversion is clean. But a Z-score of 3.5 during a spot ETF approval or a major exchange collapse means you're stepping in front of a freight train and are about to get liquidated. Pure math models break down when the underlying market regime completely shifts.

Was faced with a choice to either add a million hard coded if/then volatility filters (which always leads to curve fitting) or build something that actually understands the market context.

I chose the latter and bolted on a lightweight feedforward neural network just to grade the context of the trade.

Kept it simple with a standard 3-layer MLP with ReLU activations. Nothing crazy like Transformers because I needed execution latency to remain under 50ms.

The base engine still does the heavy lifting. It scans the market and generates the signal (like short BTC on a massive funding rate divergence). But before executing it passes a 24-hour lookback vector to the neural net.

Had to be incredibly careful about normalizing the inputs so the model wouldn't blow out during outliers. Features include normalized volatility (real time VIX/DVOL percentiles relative to a rolling 30 day window), funding rate velocity (is leverage piling in aggressively or slowly grinding up), and semantic sentiment score (scraping major news feeds for macro keywords).

The NN spits out a confidence score from 0.0 to 1.0. If the score is under 0.70 it kills the trade entirely. Model recognizes the regime is too toxic for mean reversion. If its over 0.70 it executes the trade but scales the position size linearly with the score.

Fundamentally changed my drawdowns. Now the bot actually understands when the market is trending heavily and stays out of the way.

Documented the whole architecture including the hyperparameter tuning and how I tied the ML layer into the quant model. Let me know if you want the link to the docs.

How are you guys surviving regime shifts? Still relying on hard coded moving average filters or transitioning to dynamic ML sizing?