[Paper Review] DeepFolio: Convolutional Neural Networks for Portfolios with Limit Order Book Data

Aiusha Sangadiev, Rodrigo Rivera-Castro+6 more|arXiv (Cornell University)|Aug 27, 2020

Financial Markets and Investment Strategies22 references4 citations

TL;DR

DeepFolio proposes a convolutional neural network (CNN)-based model for predicting price movements from limit order book (LOB) data and using these predictions to optimize dynamic investment portfolios. It outperforms state-of-the-art models like DeepLOB on both benchmark FI-2010 LOB and crypto-asset datasets, achieving superior accuracy and Sharpe ratio in portfolio allocation with rebalancing, especially under scarce data conditions and during market stress periods.

ABSTRACT

This work proposes DeepFolio, a new model for deep portfolio management based on data from limit order books (LOB). DeepFolio solves problems found in the state-of-the-art for LOB data to predict price movements. Our evaluation consists of two scenarios using a large dataset of millions of time series. The improvements deliver superior results both in cases of abundant as well as scarce data. The experiments show that DeepFolio outperforms the state-of-the-art on the benchmark FI-2010 LOB. Further, we use DeepFolio for optimal portfolio allocation of crypto-assets with rebalancing. For this purpose, we use two loss-functions - Sharpe ratio loss and minimum volatility risk. We show that DeepFolio outperforms widely used portfolio allocation techniques in the literature.

Motivation & Objective

• To address limitations in existing LOB-based price prediction models, particularly their sensitivity to data scarcity and poor generalization.
• To develop a deep learning framework that effectively captures spatiotemporal patterns in LOB data for improved price trend forecasting.
• To extend price prediction into actionable portfolio management by integrating predictions into dynamic, rebalanced investment strategies.
• To evaluate the performance of DeepFolio under diverse market conditions, including high-stochasticity crypto-asset markets and crisis periods.
• To compare the efficacy of different loss functions—Sharpe ratio and minimum volatility—for portfolio optimization using deep learning.

Proposed method

• Utilizes a 1D convolutional neural network (CNN) architecture to process raw limit order book data, extracting hierarchical price and volume patterns across bid and ask levels.
• Processes LOB data as time-series tensors representing bid and ask queues at multiple price levels, enabling end-to-end learning of price movement signals.
• Employs transfer learning by pre-training on a large, high-liquidity dataset (FI-2010) and fine-tuning on smaller, crypto-asset datasets to improve generalization under data scarcity.
• Integrates price predictions into portfolio optimization via two distinct loss functions: Sharpe ratio maximization and minimum volatility risk minimization.
• Applies a rebalancing strategy every 50 minutes, updating portfolio weights based on predicted returns using a multiplicative return update rule: $ p_t = p_{t-1} \cdot \frac{r_t}{r_{t-1}} \cdot w_{t-1} $.
• Trains the model using cross-entropy loss for classification of price direction (up/down) and fine-tunes the final layers with risk-adjusted objectives for portfolio performance.

Experimental results

Research questions

• RQ1 Can a deep CNN-based model like DeepFolio outperform existing state-of-the-art models such as DeepLOB in predicting price movements from LOB data, especially under data-scarce conditions?
• RQ2 To what extent can transfer learning improve generalization of LOB-based models when applied to low-liquidity, high-stochasticity crypto-asset markets?
• RQ3 Does integrating deep learning predictions into portfolio optimization with risk-adjusted loss functions (e.g., Sharpe ratio) lead to superior portfolio performance compared to classical methods?
• RQ4 How does DeepFolio perform during periods of market stress, such as the COVID-19 crisis, compared to baseline portfolio strategies?
• RQ5 Which loss function—Sharpe ratio or minimum volatility—yields better risk-adjusted returns when used to train the portfolio allocation policy in DeepFolio?

Key findings

• DeepFolio achieves 89.9% accuracy, 93.14% precision, 89.9% recall, and 90.4% F1-score at a 1-minute prediction horizon on the FI-2010 benchmark, outperforming DeepLOB’s 86.69% accuracy and 87.44% F1-score.
• On the crypto-asset dataset, DeepFolio improves F1-score by 4.56 percentage points (from 57.6% to 61.57%) at a 20-minute prediction horizon, demonstrating robustness under data scarcity.
• Transfer learning enables DeepFolio to generalize effectively: performance on unseen crypto-assets remains strong, indicating learning of general LOB patterns rather than overfitting.
• DeepFolio with Sharpe ratio loss achieves the highest Sharpe ratio (0.053069) and mean return (0.040126) on the test set, outperforming Markowitz (SR: 0.025998), 1/n (0.040234), and DeepFolio with MV loss (0.041746).
• In cumulative log-return analysis, DeepFolio with Sharpe ratio loss shows the strongest performance during the COVID-19 crisis period (starting February 2020), confirming resilience under market stress.
• The model maintains strong performance across all prediction horizons (k=1 to k=20), with consistent improvements over baselines, especially in recall and F1-score at longer horizons.