cerne.pro
Message me on LinkedIn
EN·PT
← Work
data · finance2026 · Co-author (with Ana Beatriz)

NeuraTrade

Unsupervised anomaly detection over 14 years of B3 equities with an LSTM-autoencoder, built so the methodology actually holds up.

Most deep-learning anomaly-detection work targets US markets and crypto. NeuraTrade aims it at Brazilian equities — PETR4, VALE3, AMER3, ITUB4 — training only on 2010–2019 "normality" and flagging the years after by reconstruction error, then cross-checking each hit against a timeline of real Brazilian market events.

The point wasn't a bigger model; it was a defensible one. Switching the error aggregation to max doubled recall (0.16 → 0.35) and raised precision (0.55 → 0.84). A conditional autoencoder conditioned on USD/BRL and the VIX separates systemic shocks (COVID-2020) from idiosyncratic ones (the Americanas collapse, 2023). And where ROC-AUC flattered the model at 0.84, PR-AUC exposed the real 0.15 on a rare class — so PR-AUC became the reported metric. Reproducible offline, fixed seed, one command.

PythonTensorFlow/Keraspandasscikit-learnpytest
See the code
// get in touch

Building something where data, finance, and reliability matter? Let's talk.

Message me on LinkedInGitHubmiguelcerne.dev@gmail.com
Miguel Cerne — Software engineer for data & finance systems