Table of contents

  1. System Architecture: How Aerondight Scores Stocks
  2. Layer 1: Regime Classifier
  3. Layer 2: Dual Scoring Models
  4. Layer 3: Signal Output
System Architecture: How Aerondight Scores Stocks System Architecture: How Aerondight Scores Stocks

System Architecture: How Aerondight Scores Stocks

12/31/2024 / 3 minutes to read / Tags: methodology

Aerondight Systems is built as a three-layer pipeline. Each layer has a distinct job, and the output of one feeds into the next.

Layer 1: Regime Classifier

At the top sits a Hidden Markov Model (HMM) paired with an XGBoost validator. Together, they classify the current market environment into one of four states:

  • Bull Broad — real economy rallying, broad participation, low volatility
  • Bull Tech — tech-led narrow rally, weak breadth
  • Correction — grinding bear, elevated volatility
  • Crisis — sharp crash, extreme volatility, rare but devastating

When both models agree on the regime, signal confidence increases. The regime label feeds directly into the scoring engine — the same stock can score differently in a broad bull market versus a grinding correction.

Layer 2: Dual Scoring Models

Two parallel models score every stock in the universe (~900 S&P 500 + S&P 400 MidCap names):

Swing Trade Model — 1-3 month horizon. Optimized for capturing intermediate moves. Multi-factor scoring across fundamental, technical, and sector dimensions.

Long Term Model — 1+ year horizon. Built for patient capital. Same factor categories, different weights and sensitivity to regime shifts.

Each stock receives a conviction score (1-10) and a clear signal: BUY, WATCH, or SELL.

Layer 3: Signal Output

The final layer synthesizes both models. When the swing trade and long-term models both agree on BUY, the system outputs its highest conviction signal. Signal transitions matter more than static signals — a stock moving from WATCH to BUY is an entry trigger.

graph TD
    A["REGIME CLASSIFIER\nHMM + XGBoost Validation\n4 states"] --> B
    A --> C
    B["SWING TRADE MODEL\n1-3 month horizon"] --> D
    C["LONG TERM MODEL\n1+ year horizon"] --> D
    D["SIGNAL OUTPUT\nBUY / WATCH / SELL + conviction score"]

This architecture keeps each component focused and testable. The regime classifier doesn’t know about individual stocks. The scoring models don’t decide market conditions. And the signal output simply asks: do these independent assessments agree?

← Back to research