Machine Learning Portfolios
Systemic Alpha
+
Human Judgement
No Compromises
Blackridge Capital deploys two machine learning engines - a Relative Strength scorer and a Reinforcement Learning system - to construct consistently outperforming long/short equity portfolios with over a decade of validated market history. Every position is reviewed and monitored by human traders.
Years Backtested
10+
Stocks Scored Daily
1500+
RL Models Running
39
GICS Sectors
11
Two models. One edge.
Our ML infrastructure pairs a rules-driven relative strength layer with an iteration trained reinforcement learning signal engine. Each doing the work it does best, monitored at every step by experience traders.
Model I - RS System
Relative Strength
Portfolio Engine
Every trading day, our RS scorer evaluates over 1500 equities against their GICS sector benchmarks, computing z-score deviations across multiple return windows. Stocks are ranked at their extremes of sector relative momentum to generate long and short trade candidates with statistically meaningful differentiation from the noise of market volatility. The model learns the difference between a stock thats just rising with the market tides and one that's got meaningful buyers momentum to outperform its peer universe.
-
Universe - 1,500+ equities daily
-
Benchmark - 11 GICS sector ETFs
-
Signal Type - Z score rank vs sector
-
Portfolio Size - 20 to 35 long & short
-
Refresh - Daily scoring, continuous ranking
Model II - RL Signal Engine
Reinforcement Learning
Signal Engine
Our RL engine runs 39 Proximal Policy Optimization (PPO) trading models in daily inference, each trained to identify market movement patterns that precede outsized price movements. Unlike traditional supervised models that learn from labeled outcomes, PPO agents learn through trial & error, developing market trading intuitions that mirror the actions of an experienced discretionary trader, scaled across dozens of sector specific markets simultaneously. Signals are classified within a four state regime framework (trending, mean-reverting, volatile & transitional) to determine the best action.
-
Architecture - (PPO) Proximal Policy Optimization
-
Active Models - 39 parallel inference agents
-
Regimes - 4 state classification (R1 - R4)
-
Optimal Hold - typically 10 to 20 trading days
-
Refresh - daily cron inference
How a trade is built
From signal to portfolio position
1. Scoring
The RS engine scores 1,500+ stocks before market open, ranking each against its sector ETF benchmark using z-score deviation across return windows.
2. RL Layer
High ranking RS candidates are cross referenced against active RL signals. PPO agents assess whether market state conditions support the momentum thesis implied by RS ranking.
3. Human Review
No position enters the portfolio without human trader review. Our traders evaluate macro context, news catalysts and risk concentration before ML signals become executed.
4. Manage Position
Positions are monitored against regime shifts, sector breadth signals and volatility thresholds. Exits are systematic and disciplined.