Unveiling the Power of Machine Learning in Portfolio Management

4 min readJan 18, 2024


In the dynamic world of portfolio management, machine learning stands as a beacon of innovation, promising to refine investment strategies with its computational prowess. Our latest analysis showcases the extraordinary potential of incorporating machine learning factors, particularly evident in portfolios that blends an AI view over Bitwise weekly allocations.

Live Performance Analysis: A Testament to Strategy

The first graph presents a live performance analysis, emphasising the strategy’s remarkable success in real-time application. Our active trading approach, rooted in a sophisticated ranking model developed via XGBoost, has yielded a blend of high returns and controlled risk.

This method underscores the potential of machine learning in navigating the complexities of the crypto markets.

ML Portfolio PnL vs Bitwise vs Bitcoin Market from April to December 2023 (Live Execution)
  • Risk: A balanced figure of 36.7% showing a considered approach to volatility.
  • Current Allocation: Maintained at 100%, indicating full deployment of capital.

Live Allocation: The Real-World Application

When theory meets practice, we witness the true capacity of our strategy. The live allocation graph provides evidence of our model’s adaptability and responsiveness to market movements, reinforcing the value of machine learning in portfolio management.

Crypto Asset Allocations, weekly rebalanced (from April to December 2023)

Backtest Brilliance: Outmanoeuvring Bitwise Allocations

Our backtesting efforts were not merely an academic exercise but a strategic endeavour aimed at outperforming the established Bitwise allocations. The focus on long positions has been a calculated choice, informed by the potential for growth identified by our machine learning model.

Combinatorially Purged Cross Validation Backtest of the ML portfolio model vs Bitwise

ML Model Key Insights

With only an unweighted precision of merely 0.52, our portfolio has achieved an impressive outcome in the past six months, demonstrating effectiveness in predicting single asset returns.

ROC curve (receiver operating characteristic curve) of the long/short ML model model at all classification thresholds

Accuracy: Achieving a score of 0.56 indicates a reliable prediction model.

Precision: Both weighted and unweighted precision scores reveal the model’s ability to identify positive returns effectively.

ML model metrics summarized:

  • Accuracy: 0.56 ±0.011 0.561
  • Precision: 0.519 ±0.005 0.508
  • Precision (wgt): 0.579 ±0.011 0.561
  • NPV: 0.519 ±0.009
  • NPV (wgt): 0.495 ±0.023

The Magic of low correlation in Weekly Returns

The moderate correlation coefficient of 0.596 suggests diversification benefits within the portfolio. The strong monthly correlation coupled with the low weekly correlation leads the model to identify short-term mean reversion effects to market trend and double momentum effects in single stock prices.

Monthly vs Weekly portfolio returns correlation analysis

Shapley Value Analysis

The Shapley value graph offers a window into the contribution of each feature within our global prediction model, highlighting the impact of each factor on the overall performance.

ROC 1 and SMA 2 spread are the indicators primarily responsible for strong upward movements, especially when they take on low values within their range, while SMA 4 spread is most significant in identifying future downward movements, particularly when it approaches a value near zero.

Ranked model features by Shapley value importance

The Non-Linear Dance of SMA Factors

Delving into the intricacies of our strategy, we encounter the fascinating interplay of non-linear factors that contribute to superior performance.

The analysis extends to the exploration of Simple Moving Average (SMA) factors across various weekly windows. It is this non-linear relationship among these factors that has unlocked additional performance gains, a discovery that emphasizes the nuance required in financial modeling.

Shapley value dependency graph of SMA 8 spread and SMA 2 spread

The non-linearity of the relationships between various indicators also serves to confirm or refute their behavior: in this case, the negative prediction on BTC of May 14, 2023, is partly contradicted by the SMA 8 spread indicator, while it is confirmed by all others. Indicators close to zero on metrics representing the distance from the moving average exert a decidedly strong effect regarding the prediction, but they display opposite behaviors over different timeframes.

Effect of non-linearity in probability output for one of the portfolio assets

Embracing the Future of Investing

As we stand on the cusp of a new era in investing, our journey with machine learning-enhanced portfolios presents a compelling case for the integration of advanced analytics in investment strategies. The synergy of machine learning and financial expertise heralds a future where data-driven decisions pave the way for robust and resilient portfolio management.