Source-linked AI summary
Classification-based Financial Markets Prediction using Deep Neural Networks
Matthew Dixon, Diego Klabjan, Jin Hoon Bang
TL;DR
The paper addresses limited prior application of DNNs to algorithmic trading by applying them to financial market direction classification. It trains a single model across many instruments and signals, then evaluates a simple trading strategy, reporting 42% average accuracy and sample mean Annualized Sharpe Ratios as high as 3.29. The authors conclude that concurrently trained DNN classifiers have substantial predictive capabilities on the studied CME futures data.
Problem
DNN application to algorithmic trading had not been previously researched, partly because computational complexity hindered its use.
Method
The paper trains DNN classifiers on a single model combining signals across 43 CME commodity and FX futures, using deep feed-forward learning and accelerated computation.
Results
42% average directional accuracy and sample mean Annualized Sharpe Ratios as high as 3.29 are reported across the studied futures and simple trading strategy.
Takeaways & Limitations
DNNs have substantial predictive capabilities as classifiers when trained concurrently across several markets on labelled CME futures data.
Takeaways & Limitations
The backtesting assumptions, including trade execution assumptions and no live simulation, are inadequate to demonstrate alpha-generation capabilities.
Abstract
from arXiv · showhide
Deep neural networks (DNNs) are powerful types of artificial neural networks (ANNs) that use several hidden layers. They have recently gained considerable attention in the speech transcription and image recognition community (Krizhevsky et al., 2012) for their superior predictive properties including robustness to overfitting. However their application to algorithmic trading has not been previously researched, partly because of their computational complexity. This paper describes the application of DNNs to predicting financial market movement directions. In particular we describe the configuration and training approach and then demonstrate their application to backtesting a simple trading strategy over 43 different Commodity and FX future mid-prices at 5-minute intervals. All results in this paper are generated using a C++ implementation on the Intel Xeon Phi co-processor which is 11.4x faster than the serial version and a Python strategy backtesting environment both of which are available as open source code written by the authors.
1 Introduction
The paper applies deep neural networks to classify financial market movement directions, addressing historical ANN limitations through richer multi-instrument data and accelerated training. It contrasts classification outputs with regression approaches and evaluates the resulting model through financial-market backtesting.
- Motivation and contribution: DNNs are presented as a way to model complex nonlinear relationships while reducing ANNs’ tendency to overfit, using many-core hardware to train and tune model parameters.The paper identifies nonstationarity, nonlinearity, noisiness, overfitting, convergence problems, and computational difficulty as relevant challenges.
- Training and implementation: The study uses feed-forward DNNs trained with back-propagation and gradient descent, while mini-batching and matrix formulations support efficient Intel Xeon Phi execution.The paper presents these choices as responses to the computational expense of training DNNs and sweeping their parameters.
- Motivation and contribution: The main contribution is applying DNNs to financial time-series data to classify market movement directions rather than predict price levels.This contrasts with traditional level-based methods trained separately for individual instruments.
- Data and representation: A single model combines 43 commodity and FX futures with lags, moving averages, and moving correlations to capture memory and cross-symbol co-movements.The resulting feature set is intended to provide a richer dataset for exploring complex patterns.
- Classification design: The model outputs −1, 0, or 1 for negative, flat, or positive movement, with a 1×10−3 threshold selected to balance class labels.The discrete output is intended for simple strategies that depend on forecast direction rather than magnitude.
2 Deep Neural Network Classifiers
The paper formulates a fully connected feed-forward DNN classifier that outputs probabilistic market states and trains its weights by minimizing cross-entropy. Mini-batching recasts feed-forward and back-propagation computations as matrix operations for efficient optimization.
- Network architecture: Each observation is represented by an input feature vector whose nodes connect fully to the next network layer through learned weights.The network contains multiple hidden layers between input and output layers.
- Training objective: Training minimizes a cross-entropy objective over the network weights, using sigmoid activations in hidden layers and analytic derivatives for back-propagation.The supplied passages identify cross-entropy as the optimization objective and the sigmoid derivative used in training.
- Output representation: The classifier uses one-of-k target and output encodings, with softmax probabilities assigning a normalized weighting to each output state.The softmax output ensures that state probabilities sum to unity.
- Optimization: Stochastic gradient descent updates each layer’s parameter vector after randomly sampling an observation, with γ specifying the learning rate.The implementation notes that SGD is used for the highly non-convex utility function.
- Mini-batching: Mini-batching processes b observations together, adds a batch dimension to intermediate variables, and expresses layer computations and weight updates as matrix-matrix products.This formulation is intended to improve computational performance while leaving the dimensions of the weights unchanged.
3 The Data
The dataset contains normalized five-minute mid-prices and engineered features for 43 CME commodity and FX futures. Training uses consecutive observations followed by a temporally subsequent test set, with rolling evaluation described for later experiments.
- Dataset: The historical dataset contains five-minute mid-prices for 43 CME-listed commodity and FX futures from March 31, 1991 to September 30, 2014.The most recent fifteen years are used because earlier data are less liquid for some symbols.
- Preprocessing: Each feature is normalized by subtracting its mean and dividing by its standard deviation.This normalization is applied before forming the training and test sets.
- Data split: The training set contains 25,000 consecutive observations, followed by a test set of the next 12,500 observations.The sets are rolled forward ten times in the later application described by the paper.
- Feature engineering: The aggregate training dataset contains 9,895 features combining price differences, lags, moving averages, and pair-wise return correlations across symbols.These features are intended to capture historical memory and co-movements between instruments.
4 Implementation
The implementation uses a deep, fully connected network with heuristic learning-rate control and mini-batch training, optimized for parallel execution on an Intel Xeon Phi.
- Network architecture: The network contains five learned fully connected layers, tapering hidden layers from 1000 neurons and producing 129 outputs for 43 futures.The architecture contains 12,174,500 weights in total.
- Training procedure: The learning rate is swept from [0.1, 1] in increments of 0.1 and halved when cross-entropy fails to decrease between epochs.The same learning rate is used for all layers.
- Parallel implementation: Mini-batching expresses back-propagation in matrix form, enabling hardware-optimized numerical linear algebra routines for the Xeon Phi.The method generates epoch subsets and mini-batches before iterating through network layers.
- Performance: 11.4x faster than the serial algorithm, training the DNN on the described data takes approximately 8 hours on an Intel Xeon Phi.The timing includes test-set error calculations and supports overnight batch retraining.
5 Results
The results evaluate DNN classification across 43 CME commodity and FX futures using rolling test windows and ten walk-forward experiments. Average accuracy concentrates near 0.35, while performance varies substantially across instruments, with some performing no better than white noise.
- Classification accuracy: Ten classification errors are calculated from 130-day moving test windows for each of the 43 futures.Figure 2 summarizes the resulting classification accuracies with box-and-whisker plots for each symbol.
- Classification accuracy: 0.35 is the accuracy around which the average classification distribution has heavier density, slightly better than random selection.Figure 3 summarizes average accuracy over ten samples across the 43 futures.
- Instrument comparisons: Table 1 reports the top five instruments by mean classification rate, alongside F1-scores and aggregate means and standard deviations across all 43 futures.F1-scores are included because they are less sensitive to class imbalance than classification accuracy.
- Instrument comparisons: The worst five instruments perform no better than, or worse than, white noise on average over the ten experiments.This contrasts with the separate identification of the highest-performing instruments.
6 Strategy Backtesting
The paper backtests a simple classification-driven trading strategy using walk-forward optimization across 43 CME commodity and FX futures. Performance is evaluated through returns, profit, drawdown, Sharpe ratios, and benchmark correlations under simplified execution assumptions.
- Walk-forward optimization: The walk-forward procedure trains on 25,000 observations, evaluates 12,500 out-of-sample observations, then repeats after sliding the window forward by 1,000 observations for ten windows.The model is trained using all symbol data and engineered time series, while cumulative P&L is calculated separately for each symbol.
- Trading strategy: The strategy buys, holds, or sells one lot according to whether the predicted next movement is positive, neutral, or negative.A positive label closes shorts and opens longs, a zero label holds, and a negative label closes longs and opens shorts.
- Backtest assumptions: The backtest assumes immediate mid-price fills, ignored transaction costs and slippage, unrestricted holding periods, and no operational risk controls.It also assumes sufficient surplus cash for margin, zero interest on margin, and no material market impact from one-lot orders.
- Backtest assumptions: The authors state that these execution assumptions and the absence of live simulation are inadequate for demonstrating alpha generation, but useful as a starting point for commercial application.This explicitly limits how the backtest should be interpreted.
- Performance measures: Figure 5 summarizes time-averaged daily returns for each of the 43 futures over 130-day horizons, with red bordered squares marking sample averages.Returns are aggregated from intraday P&L changes and annualized.
- Performance measures: Figure 6 compares cumulative unrealized net dollar profit under perfect foresight and DNN prediction for one 130-day front-month Platinum horizon.The green line denotes perfect foresight and the blue line denotes prediction.
- Performance measures: Maximum drawdown is summarized across ten walk-forward experiments for each symbol, while Sharpe ratios are examined for the top five performing futures.The Sharpe-ratio analysis uses 12,500-observation moving periods; no benchmark is used in its calculation.
- Benchmark comparison: Strategy returns for the five most liquid instruments are weakly correlated with relevant ETF benchmarks, with absolute correlations below 0.5 in every experiment.The correlations are summarized over ten experiments.
7 Conclusion
The paper presents DNNs as classifiers of financial market movement directions using multivariate futures data and engineered temporal and cross-market features. It reports predictive capabilities and demonstrates their use in backtesting a simple trading strategy.
- Conclusion: DNNs are applied concurrently across multiple CME futures using historical prices, lags, filters, and other engineered features to classify market movement directions.The authors describe this as capturing memory and co-movements across symbols.
- Conclusion: The implementation combines DNN training with a backtesting environment to examine prediction accuracy and its relation to strategy profitability.The computational implementation uses C++ on an Intel Xeon Phi co-processor and a Python backtesting environment.
- Conclusion: The Intel Xeon Phi implementation is reported as 11.4x faster than the serial version, and the authors provide both implementations as open-source code.The reported speedup applies to the C++ implementation on the co-processor.