Strategy Guides

How Risk Radar Works: The ML Model Behind Our Strongest Signal

Risk Radar identifies stocks most likely to decline in the near term. Here's how the LightGBM model works, what it looks at, and why it's better at finding losers than winners.

Steffen FonvigMarch 24, 20263 min read

Risk Radar Is Our Best Feature — Here's Why

Most AI stock platforms market their ability to find winners. We lead with our ability to find losers. That is not modesty — it is what the data shows.

In walk-forward backtesting from 2023 to 2026, our LightGBM model's strongest and most consistent signal is identifying stocks that will underperform. Bottom-ranked stocks declined an average of -3.2% over 5 trading days, and the model correctly ranked outcomes 81% of the time.

The Technical Stack

50 Engineered Features

Every trading day, for every stock in our universe of 7,700+, we compute:

  • Momentum features: RSI, MACD histogram, rate of change, Williams %R
  • Trend features: SMA 50/200 alignment, ADX strength, MACD crossover signals
  • Volatility features: Bollinger Band position, ATR, historical volatility
  • Volume features: On-balance volume, volume ratio on up vs down days
  • Price action: Multi-timeframe returns (1d, 5d, 20d), distance from 52-week high
  • These raw indicators are then engineered into 50 ML features that the model consumes.

    Optuna-Optimized LightGBM

    We use LightGBM (gradient boosting) with hyperparameters optimized by Optuna across hundreds of trials. The model is retrained monthly using walk-forward validation — meaning it never sees future data during training.

    Two models work together:

  • Return Regressor — Predicts the 5-day forward return
  • Direction Classifier — Predicts the probability of a positive return
  • The ML score blends both: 60% return percentile + 40% direction probability.

    Market Regime Awareness

    The model adjusts based on whether we are in a BULL, BEAR, or CHOPPY market (detected automatically using SPY data). In bear markets, the model becomes more cautious — tightening position sizing and raising signal thresholds.

    How to Use Risk Radar

    The simplest workflow takes 30 seconds every morning:

  • Open the Risk Radar page
  • Cross-reference against your current holdings
  • If any of your stocks appear, investigate further
  • Consider tightening stop-losses or reducing position size
  • Risk Radar is not telling you to panic-sell. It is flagging elevated risk so you can make informed decisions.

    Important Caveats

  • All performance numbers are from backtesting (2023-2026). Live tracking began March 2026.
  • No model is perfect. The 81% ranking accuracy means it is wrong about 1 in 5 times.
  • Risk Radar should be one input in your process, not the only one.
  • Past performance does not guarantee future results.

  • See today's Risk Radar →

    risk radarmachine learningLightGBMmethodology
    SF

    Steffen Fonvig

    Founder & CEO, Fonvig Group

    Entrepreneur and founder building companies across fintech, media, and health tech since 2013. Creator of Fillipio, an AI-powered stock screening platform. Based in Oslo, Norway.