This version: December 12, 2013
Applying Deep Learning to Enhance Momentum Trading Strategies in Stocks
Lawrence Takeuchi * Yu-Ying (Albert) Lee
Abstract
We use an autoencoder composed of stacked restricted Boltzmann machines to extract features from the history of individual stock prices. Our model is able to discover an en- hanced version of the momentum effect in stocks without extensive hand-engineering of input features and deliver an annualized re- turn of 45.93% over the 1990-2009 test period versus 10.53% for basic momentum.
1. Introduction
Price momentum is the empirical finding (Jegadeesh & Titman, 1993) that stocks with high past returns over 3-to-12 months (winners) continue to perform well over the next few months relative to stocks with low past returns (losers). Subsequent studies find that this mo- mentum effect continues to remain robust in the US after becoming widely known and applies to interna- tional stocks as well as other asset classes including for- eign exchange, commodities, and bonds (Asness et al., 2013). For finance academics the fact that a simple strategy of buying winners and selling losers can ap- parently be profitable challenges the notion that mar- kets quickly incorporate available information into as- set prices. Indeed, Fama and French (2008) describe momentum as the “premier anomaly” in stock returns.
The momentum trading strategy, along with its many refinements, is largely the product of a vast, ongoing effort by finance academics and practitioners to hand- engineer features from historical stock prices. Recent advances in deep learning hold the promise of allowing machine learning algorithms to extract discriminative information from data without such labor-intensive feature engineering and have been successfully applied to fields such as speech recognition, image recognition,
*Corresponding author.
ltakeuch@stanford.edu yy.albert.lee@gmail.com
and natural language processing (Bengio et al., 2012).
In this paper we examine whether deep learning tech- niques can discover features in the time series of stock prices that can successfully predict future returns. The objective is a challenging one. While most research in deep learning considers tasks that are easy for humans to accomplish, predicting stock returns using publicly available information is notoriously difficult even for professional investors given the high level of noise in stock price movements. Furthermore, any patterns that exist are subject to change as investors themselves learn over time and compete for trading profits.
Considering the pervasive use of historical price charts by investors and noting that the primary mode of anal- ysis is visual, we take an approach similar to that used by Hinton and Salakhutdinov (2006) to classify hand- written digits. In particular, we use an autoencoder composed of stacked restricted Boltzmann machines (RBMs) to extract features from stock prices, which we then pass to a feedforward neural network (FFNN) classifier.
2. Data
We obtain data on individual US stocks from the Cen- ter for Research in Security Prices.
2.1. Sample selection
We restrict our analysis to ordinary shares trading on NYSE, AMEX, or Nasdaq. To mitigate the impact of any market microstructure-related noise, we exclude stocks with monthly closing prices below $5 per share at the time of portfolio formation. This step also re- duces the number of examples with extreme returns.
The training set covers the period from January 1965 to December 1989, which coincides with the period ex- amined by Jegadeesh and Titman (1993), and contains 848,000 stock/month examples. The test set covers the period from January 1990 to December 2009 and contains 924,300 stock/month examples. On average
Applying Deep Learning to Enhance Momentum Trading Strategies in Stocks
there are 3,282 stocks in the sample each month.
2.2. Input variables and preprocessing
We want to provide our model with information that would be available from the historical price chart for each stock and let it extract useful features without the need for extensive feature engineering. For every month t, we use the 12 monthly returns for month t−2 through t − 13 and the 20 daily returns approximately corresponding to month t.1 We also use an indicator variable if the holding period, month t + 1, falls in January.2 Thus we have a total of 33 input variables for each stock/month example.
Next we compute a series of 12 cumulative returns using the monthly returns and 20 cumulative returns using the the daily returns. We note that price mo- mentum is a cross-sectional phenomenon with winners having high past returns and losers having low past re- turns relative to other stocks. Thus we normalize each of the cumulative returns by calculating the z-score rel- ative to the cross-section of all stocks for each month or day. Figure 1 illustrates preprocessing pipeline us- ing the 12 monthly returns for one example in the test set.
Finally we use returns over the subsequent month, t + 1, to label the examples with returns below the median as belonging to class 1 and those with returns above the median to class 2.
3. Deep Learning Model
3.1. Model and learning algorithm
We follow an approach similar to that introduced by Hinton and Salakhutdinov (2006) to train networks with multiple hidden layers. Our model consists of a stack of RBMs, which unlike full Boltzmann machines have no intra-layer connections. Each RBM consists of one layer of visible units (the inputs) and one layer of hidden units connected by symmetric links. The output of each RBM serves as the input to the next RBM in the stack.
We train the encoder network (see Figure 2) layer by layer in a pretraining step. Following Hinton (2010), we split the dataset into smaller, non-overlapping mini-batches. The RBMs in the encoder are then un- rolled to form an encoder-decoder, which is fine tuned using backpropagation. In our implementation, the number of hidden units in the final layer of the encoder is sharply reduced, which forces a reduction in dimen- sionality. As described below, however, the size of this bottleneck layer is an outcome of the optimization pro- cedure we use to specify the network dimensions rather than an explicit design choice.3
At this stage, the encoder outputs a low dimensional representation of the inputs. The intention is that it retains interesting features from the historical stock chart that are useful for forecasting returns, but elim- inates irrelevant noise. We use the weights estimated thus far to initialize the corresponding weights in the full network, which is composed of the encoder and a FFNN classifier. The final step is to train the entire network using the labeled examples via backpropaga- tion.4
3.2. Network specification
We use hold-out cross validation to determine the number of layers and number of units per layer in our network. In particular, we further divide the train- ing set into two subsets covering 1965-1982 and 1983- 1989, respectively. Each model specification is trained on the first subset and then tested on the second. We choose this approach over k-fold cross validation since we have a sufficiently large dataset and more impor- tantly want to avoid the look-ahead bias that could arise from training the model with data not available at a given historical date.
3Hinton and Salakhutdinov (2006), by contrast, omit an explicit bottleneck layer in their digit classification example and use a 784-500-500-2000-10 network.
4See Ng et al. (2013) for a tutorial on training deep networks.
Figure 1. Preprocessing of inputs
1We choose these inputs given the casual observation that investors tend to use higher frequency prices when viewing charts covering shorter windows of time.
2Turn-of-the-year patterns in stock returns were well- known prior to Jegadeesh and Titman (1993). In any case, using dummy variables for each calendar month gives sim- ilar results.
Applying Deep Learning to Enhance Momentum Trading Strategies in Stocks
Figure 2. Network Architecture
To keep the task manageable, we fix the number of units in the penultimate hidden layer at 50.5 Thus, our first set of candidate specifications is 33−s2 −50−2 and the second is 33−s2 −s3 −50−2 where sl de- notes the number of units in layer l. A grid search over the number of units using our hold-out cross val- idation scheme finds that s2 = 40 and s3 = 4 give the lowest classification error. The next set of candi- date specifications is 33−s2 −s3 −s4 −50−2. Given the large number of dimensions, we use s2 = 40 and s4 = 4 and only search over s3. We find that adding another layer does not reduce the classification error.6 Thus our final specification is the five-layer network 33−40−4−50−2 consisting of an encoder that takes 33 inputs and reduces them to a 4 dimensional code and a classifier that takes these 4 inputs and outputs the probabilities for the two classes. While the ap- proach taken is admittedly heuristic, we believe that it provides a disciplined method to specify a base case model.
4. Results
We train our model using the examples from 1965-1989 and test it using those from 1990-2009. In this section we keep both the network configuration and weights fixed, but discuss later how these could be updated over time.
5However, once we determine the specification of the stacked autoencoders we verify that this is a reasonable choice.
6Our ongoing work suggests that additional layers may be useful when the number of features are increased.
Table 1. Confusion Matrix Predicted
12 Actual 1 22.38% 27.45%
2 19.19% 30.97%
4.1. Classification performance
Table 1 summarizes the results in a confusion matrix with the entries scaled by the total number of test examples. The model achieves an overall accuracy rate of 53.36%. The model is correct 53.84% of the time when it predicts class 1 and a somewhat lower 53.01% of the time when it predicts class 2.
These probabilities alone, however, do not provide a complete picture of the model’s performance since in- vestors actually care more about returns in their objec- tive function. In particular, we would like to compare the realized returns for the stocks predicted to be in class 1 against those for stocks predicted to be in class 2. Pooling all months in the test set, the average one- month holding period return is 0.11% for stocks pre- dicted to be in class 1 and 1.50% for those predicted to be in class 2, or a difference of 1.39%. Alternatively, we can use past 12 month returns (from month t − 13 to t − 2) and predict that examples with past returns below the median will be in class 1 and those above the median in class 2. This basic momentum strategy produces an average holding period return of 0.64% for stocks in class 1 and 1.20% for those in class 2, or a difference of 0.55%.
4.2. Information content of class probabilities
While these results appear promising, we have used only a portion of the information produced by the model. Figure 3 shows the relation between the es- timated probability of being in class 2 according to the model versus the holding period return that is ac- tually realized for the test set, where we use a Gaus- sian kernel regression to smooth the curve. We see an increasing relation indicating that a higher class 2 probability leads to higher realized returns on average. Since there is no requirement to hold every stock, an investor could clearly do better if he bought and sold stocks in the tails of the distribution rather than using the 50% threshold to form long and short portfolios.
We rank all stocks each month by their class 2 proba- bilities and buy those in the top decile and sell those in the bottom decile. Next month we close out these posi- tions and form new long and short portfolios. Repeat-
Applying Deep Learning to Enhance Momentum Trading Strategies in Stocks
4 10
3 10
2 10
1 10
0
10
89 90 91 92 93 94 95 96 97 98 99 00 01 02 03 04 05 06 07 08 09 10
Enhanced strategy
Basic strategy
0.020 0.015 0.010 0.005
0 −0.005 −0.010 −0.015 −0.020 −0.025
0.20 0.25
0.30 0.35 0.40 0.45 0.50 0.55
CLASS 2 PROBABILITY
0.60 0.65
Figure 3. Holding period returns by class 2 probability
Table 2. Average monthly momentum returns, 1990-2009.
Figure 4. Log Growth in Cumulative Wealth
strategies. We focus on these two features since they are highlighted in the finance literature (Jegadeesh & Titman, 1993) as being potential predictors of stock re- turns. By construction, the basic momentum strategy takes very extreme positions based on past 12 month returns with a 10−1 spread in this feature of more than 3 standard deviations. The enhanced strategy has a 10 − 1 spread of 0.62 showing that it is also making a bet based on past 12 month returns, but in a less ex- treme way than basic strategy. The lower half of the table shows that the enhanced strategy buys stocks that have had poor recent returns and sells those with high recent returns, consistent with the short-term re- versal effect (Jegadeesh, 1990). The basic momentum strategy, however, has similar past 20 day returns in deciles 1 and 10.
While we highlight these two features, we emphasize that the model is likely picking up other more subtle patterns in the historical price chart that are less in- tuitive to interpret. Table 3 shows that the enhanced strategy does not take overly extreme positions based on past 12 month and past 20 day returns as would be the case if we double sort on these features to form portfolios as is common practice in finance studies. We regard this finding, together with the size of the in- vestment returns, as encouraging for deep learning as it suggests that our model is not merely rediscovering known patterns in stock prices, but going beyond what humans have been able to achieve.
5. Discussion and Further Work
This study represents one of the first, as far as we are aware, applications of deep learning to stock trading and makes two main contributions to the applied ma- chine learning literature. First, we show that stacked autoencoders constructed from RBMs can extract use- ful features even from low signal-to-noise time series
Strategy
Enhanced
t-statistic
Basic 0.25% 1.35% 1.10% t-statistic 0.47 2.76 2.39
ing this process over the test set generates a monthly time series of investment returns. For comparison we also perform this analysis using past 12 month returns to form deciles. Table 2 presents the average returns for these two strategies, which we call enhanced and basic momentum, respectively. The enhanced strategy generates an average monthly return of 3.35% with a t-statistic of 9.26. In contrast, the basic strategy pro- duces a more modest, but still statistically significant, 1.10% per month.7
On an annualized basis, the returns from our model are a very impressive 45.93% versus 10.53% for basic momentum. Figure 4 shows the growth, on a logarith- mic scale, of $1 invested in each of the enhanced and basic strategies from 1990 to 2009. Ending wealth, be- fore considering implementation costs, is $7.41 for the basic strategy and $1,919, or 259 times higher, for the enhanced strategy.
4.3. Source of investment returns
We now show that the enhanced strategy is, in fact, a momentum strategy in the sense that the decile 10 stocks have higher past 12 month returns than the decile 1 stocks. Table 3 shows the past 12 month (month t−13 to t−2) and past 20 day (month t) re- turns expressed as z-scores for the enhanced and basic
7These returns are calculated as the arithmetic mean of the time series of monthly returns.
Decile 1
Decile 10
10 – 1
-1.02% -2.05
2.33% 7.75
3.35% 9.26
HOLDING PERIOD RETURN
Cumulative Wealth
Applying Deep Learning to Enhance Momentum Trading Strategies in Stocks
Table 3. Stock characteristics by strategy.
References
Asness, Clifford, Moskowitz, Tobias, and Pedersen, Lasse. Value and momentum everywhere. Journal of Finance, 68:929–985, 2013.
Bengio, Yoshua, Courville, Aaron, and Vincent, Pas- cal. Representation learning: a review and new per- spectives. ArXiv e-prints, 2012.
Fama, Eugene and French, Kenneth. Dissecting anomalies. Journal of Finance, 63:1653–1678, 2008.
Hinton, Geoffrey. A practical guide to training re- stricted Boltzmann machines. Technical Report MTML TR 2010-003, University of Toronto, 2010.
Hinton, Geoffrey and Salakhutdinov, Ruslan. Reduc- ing the dimensionality of data with neural networks. Science, 313:504–507, 2006.
Jegadeesh, Narasimhan. Evidence of predictable be- havior of security returns. Journal of Finance, 45: 881–898, 1990.
Jegadeesh, Narasimhan and Titman, Sheridan. Re- turns to buying winners and selling losers: implica- tions for stock market efficiency. Journal of Finance, 48:65–91, 1993.
Ng, Andrew, Ngiam, Jiquan, Foo, Chuan Yu, Mai, Yifan, and Suen, Caroline. UFLDL Tutorial, 2013. URL http://ufldl.stanford.edu/wiki/ index.php/UFLDL_Tutorial.
Past 12m ret – Enhanced
– Basic
Past 20d ret – Enhanced
– Basic
Decile 1
-0.39 -1.05
0.41 -0.06
Decile 10
0.23 2.03
-0.51 0.05
10 – 1
0.62 3.08
-0.92 0.11
data such as financial asset prices if the inputs are appropriately preprocessed. Second, we illustrate the potential for deep learning to reduce the need for ex- tensive feature engineering in an application area (fi- nancial markets) that has long been of interest to ma- chine learning researchers. Our model easily accom- modates returns of different frequencies as well as non- return data and produces investment results that ex- ceed those of most strategies in the vast finance liter- ature on momentum strategies.
In ongoing work, we are considering additional fea- tures such as industry and aggregate market returns as well as non-return data such as firm characteristics and macroeconomic indicators. An open question as we expand the number of inputs is whether separate autoencoders for various categories of features would perform better than combining all features in a single autoencoder.
We are also examining the impact of updating the weights in our network over time. Our current method- ology, in which we train the model once and then hold parameters fixed for the entire test period, is unlikely to be optimal as investor behavior as well as the institutional framework of the market change over time. Furthermore, even if the data were station- ary, we could improve performance by training with all the available data at each date. An important im- plementation issue is computational cost, especially as the number of features and depth of the network in- crease. We are exploring a parallel implementation of the learning algorithm that could be run on GPUs. This approach should lead to a substantial decrease in training time as the algorithm can take advantage of parallelization at the data-level (since it uses mini- batches) as well as at the network layer level. Alter- natively, a more straightforward approach would be to retrain the classifier each month, but update the autoencoder less frequently in order to limit computa- tional costs.