{
  "id": 8004386,
  "title": "Detectar volatilidad anómala en los mercados con un Temporal Fusion Transformer",
  "url": "https://urgent.news/2026/09/17/detectar-volatilidad-anomala-en-los-mercados-con-un-temporal-fusion",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-17T10:00:02.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/adrian_368e1d3e691afab697/detectar-volatilidad-anomala-en-los-mercados-con-un-temporal-fusion-transformer-20c2"
  },
  "original_language": "es",
  "account": "Detecting unusual market volatility with a Temporal Fusion Transformer\n\nPredicting a stock's price is nearly impossible and, frankly, not very useful. A more attainable and valuable question is: when is an asset poised to become abnormally volatile? Volatility spikes often precede news, not the other way around. This was the goal of this market sensor. The problem was to estimate the full distribution of returns for each asset and time horizon, and flag as anomalies the moments when reality deviates from the quantiles the model considers plausible. The architecture: TFT + quantile regression. I chose the Temporal Fusion Transformer because it combines three things I needed: Variable Selection Networks that learn which inputs matter at each moment, providing interpretability without sacrificing power. Multi-head attention over the temporal window, capturing long-term dependencies (a pattern from weeks ago) better than an LSTM. Quantile regression output: instead of a value, it predicts several quantiles (p10, p50, p90...). The distance between quantiles is the estimated uncertainty. When a real return falls outside the expected quantile interval, an anomaly alert is triggered. The metric that truly validates the model is calibration: if it says something happens 90% of the time, it must happen 90% of the time. This model's quantile calibration turned out to be practically perfect, meaning its uncertainty intervals are reliable, not decorative. Results: F1-score (anomaly detection): 0.82, MCC: 0.80, Quantile calibration: ≈ perfect. What I learned: Reformulating the problem—moving from predicting price to predicting uncertainty—makes it tractable and useful. And in finance, a model that knows its own limits (its calibration) is more valuable than one that gives precise and inaccurate answers. Published originally on adrianmoreno-dev.com.",
  "summary": "Predecir el precio de una acción es casi imposible y, francamente, poco útil. Hay una pregunta más alcanzable y más valiosa: ¿cuándo está a punto de volverse anormalmente inestable un activo? Los picos de volatilidad suelen preceder a las noticias, no seguirlas. Ese era el objetivo de este \"sensor de mercado\". El problema En lugar de estimar un único número (el retorno esperado), quería estimar…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}