See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/328907909

Deep Learning for Global Tactical Asset Allocation

Technical Report  in  SSRN Electronic Journal · January 2018

DOI: 10.2139/ssrn.3242432

CITATIONS READS

10 5,183

4 authors:

Gaurav Chakravorty Ankit Awasthi

Google Inc. Lovely Professional University
13 PUBLICATIONS   41 CITATIONS    67 PUBLICATIONS   811 CITATIONS   

SEE PROFILE SEE PROFILE

Brandon Da Silva Mansi Singhal

4 PUBLICATIONS   10 CITATIONS    8 PUBLICATIONS   12 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

risk based allocation View project

Dynamic Currency Hedging View project

All content following this page was uploaded by Gaurav Chakravorty on 07 March 2019.

The user has requested enhancement of the downloaded file.

 Deep Learning based Global Tactical Asset Allocation
∗1 †1 ‡2
Gaurav Chakravorty , Ankit Awasthi , Brandon Da Silva and Mansi Singhal
§1

1
Qplum
2
OPTrust

October 19, 2018

Abstract
We show how one can use deep neural networks with macro-economic data in conjunction
with price-volume data in a walk-forward setting to do tactical asset allocation. Low cost
publicly traded ETFs corresponding to major asset classes (equities, fixed income, real estate)
and geographies (US, Ex-US Developed, Emerging) are used as proxies for asset classes and
for back-testing performance. We take dropout as a Bayesian approximation to obtain predic-
tion uncertainty and show it often deviates significantly from other measures of uncertainty
such as volatility. We propose two very different ways of portfolio construction - one based
on expected returns and uncertainty and the other which obtains allocations as part of the
neural network and optimizes a custom utility function such as portfolio Sharpe Ratio. We
also find that adding a layer of error correction helps reduce drawdown significantly during
the 2008 financial crisis. Finally, we compare results to risk parity and show that the above
deep learning strategies trained in totally walkforward manner have comparable performance.

Keywords— tactical asset allocation, deep learning, walkforward, quantitative portfolio man-
agement

1 Qplum is a global investment management firm.

2 OPTrust invests and manages one of Canada’s largest pension funds and administers the OPSEU
Pension Plan, a defined benefit plan with over 92,000 members and retirees.
Qplum or OPTrust may or may not apply similar investment techniques or methods of analysis
as described herein. The views expressed here are those of the authors and not necessarily those
of Qplum or OPTrust. We would like to thank Dr Michael Steele for his valuable comments and
research assistance. Please refer to important disclosures at the end of this document.
∗ [email protected]
† [email protected]
‡ [email protected]
§ [email protected]

1
1 Introduction
While deep learning has had remarkable progress and is the state of the art in various domains
such as computer vision, natural language understanding, and speech recognition, it is safe to
say that investment management is still largely not driven by deep neural networks or machine
learning in general. One of the major reasons for this is lack of data available in finance as we
note later in our results. Moreover, owing to limited training data, poor generalization in out-of-
sample data and fundamentally low signal to noise ratio in financial time series data, deep learning
methods have not vastly outperformed other methods as witnessed in other domains. Traditionally
quantitative asset managers come up work with a large corpus of hand crafted signals and over
time, some of the signals might lose value in terms of predicting expected returns and new signals
have to be created. Deep learning models trained in a walk-forward manner try to do this in a
systematic manner where the model periodically adapts its weights to learn features as more data
becomes available. So instead of focusing on the model parameters, we iterate over the policy
through which we learn these weights. In this work, we show that the walk-forward learning of
deep neural network models can be accomplished effectively with careful choice of certain hyper-
parameters which determine how fast the model reacts to new data. Because of low signal to noise
ratio in financial time series data, point estimates of expected returns typically have significant
errors. Therefore, it is prudent to work in a Bayesian setting where expected returns are used in
conjunction with uncertainty estimates.
Most of the deep learning research in finance till date has focused on security selection or pre-
dicting returns for single stocks [Bao et al.(2017)Bao, Yue, and Rao], or a cross-section of stocks
[Abe and Nakayama(2018)]. [Alberg and Lipton(2017)] focuses on predicting fundamental data
for stocks and show that stock selection using the predicted fundamentals leads to improved per-
formance. While [Liew and Mayster(2017)] looks at predicting ETFs, they only work with price-
volume features. Partly as a result of features used, they found that the sweet spot of their
prediction accuracy was lower than what we notice in our experiments.
In their pioneering work, [Brinson et al.(1995)Brinson, Hood, and Beebower] pointed out that
asset allocation explained most of the variance in pension funds’ performance. Traditionally active
investors have focused on security selection to outperform the broad index and as a result financial
markets may be macro inefficient due to insufficient "smart money" being available to arbitrage
mispricing effects away [Blitz and van Vliet(2008)].
Tactical asset allocation refers to investment decisions which lead to allocation changes across
asset classes, geographies, sectors or industries with respect to a given static strategic alloca-
tion. We propose two different ways of portfolio construction. In the first approach we use
expected returns and uncertainty estimates to obtain allocation to different asset classes. The po-
sition sizing in our case is inspired from Kelly position sizing logic [Kelly(1956)], [Breiman(1961)],
[Chakravorty(2017)], and we add a layer of error correction. In the second approach, we take a
holistic portfolio construction approach and use the utility function - in this case portfolio Sharpe
Ratio, and directly train the feed-forward neural network with the utility function as the opti-
mization objective. The benefits of doing this are manifold. Firstly, better prediction of expected
returns does not directly translate to better utility function especially when performance met-
rics are risk-adjusted or anchored to a certain benchmark. Secondly, the walk-forward learning
procedure employed here naturally takes into consideration transaction and slippage costs result-
ing from excessive turnover. In this way, the network penalizes any behavior in the allocation
that hurts the utility function and these can be easily controlled using appropriate slippage or
transaction costs. Lastly, although in this work we have simply added a penalty for concentrated
portfolios, the same framework can easily be extended to work with a range of objectives. As
[Louton et al.(2015)Louton, McCarthy, Rush, Saraoglu, and Sosa] argue, pension investors should
have a lever on how tactical their plan allocations be and should be allowed to change based on
considerations such as manager skills and risk-return tradeoff. This holistic way of asset alloca-
tion can accommodate investor preferences such as the amount of active risk one wants over the
strategic allocation while maintaining a tactical tilt based on the model’s predictions. While the
utility function is single metric, it can be used to incorporate diverse investor preferences such as
loss aversion or active risk constraints [Singhal and Chakravorty(2018)].
Macroeconomic data has been shown to be useful in forecasting economic growth (refer to
[Estrella and Hardouvelis(1991)]) and could be useful detecting business cycle changes and pre-
dicting recession [Estrella and Mishkin(1996)]. [Kliesen(2018)] note that the yield curve helps

2
 ETF Description Expense Ratio Asset Class
BND Vanguard Total Bond Market ETF 0.05 US Fixed Income
BNDX Vanguard Total International Bond ETF 0.11 International Fixed Income
VWOB Vanguard Emerging Markets Govt Bond ETF 0.32 Emerging Fixed Income
VTI Vanguard Total Stock Market ETF 0.04 US Equities
VGK Vanguard FTSE Europe ETF 0.10 Europe Equities
VWO Vanguard FTSE Emerging Markets ETF 0.14 Emerging Equities
VNQ Vanguard Real Estate ETF 0.12 US Real Estate
VNQI Vanguard Global Ex-US Real Estate ETF 0.14 Ex-US Real Estate

Table 1: List of ETFs used, with a short description, expense ration and asset class

predict recession two to size quarters ahead when compared to other financial and macroeconomic
indicators. [Smalter Hall(2017)] shows that deep neural networks outperform benchmark models
for predicting civilian unemployment unto 4 quarters in future. The rest of the article is laid out
as follows. In the next section, we discuss the securities that are used, the selection criteria and
why ETFs are an efficient investment vehicle for the investment strategy we build. We also discuss
the different features derived from end of day price-volume and macroeconomic data which are
used as inputs to the deep learning models. Next, we discuss the model specification of the neural
network architectures we experimented with, followed by the specifics of the walk-forward training
procedure. Then, position sizing using the expected returns or utility function is discussed. This is
followed by backtesting assumption. Finally, we discuss the results of our experiments and report
the findings.

2 Data
2.1 Traded Securities
Since we are focusing on tactical asset allocation and not security selection, ETFs provide a low
cost vehicle for expressing investment decisions across a large cross section of securities without
undertaking the execution costs and risks involved. The selection of ETFs listed in table 1 was done
by taking into account the following factors to make sure that our reported results are investable
and realizable without significant costs.
• Coverage : ETFs provide coverage across a large cross-section of securities and alleviate the
need to explicitly execute the basket of securities that represents.
• Expense : The selection of ETFs was done to make sure that the expense does not lead to
a significant drag in performance. Average expense of the ETFs used in the portfolio is less
than 13 cents.
• Liquidity : Active investment strategies require trading in and out of positions frequently.
Higher liquidity leads to lower execution slippage.

2.2 Input Features

We have 46 input features - 14 based on macro-economic data and 32 based on price volume data.
• Macroeconomic : Listed in table 2. The choice of which economic indicators was mainly
driven by two ideas - access to clean data and relevance for forecasting returns and detecting
business cycle changes. It is important to note that the economic data releases are subject to
revisions and restatements and naive use of these datasets can lead to significant look-ahead
bias. Point in time estimates should be collected to ensure that the learning algorithm is not
looking ahead.
• Price-Volume : We use features which capture historical price action over different horizons.
In particular, we have 4 features per security, and these correspond to lagged returns over
[260, 130, 60, 30] days respectively. The choice of the historical returns duration is done partly
to eliminate short term noise from the daily return numbers and partly to keep turnover of
the resulting investment strategy low.

3
 10 Year - 2 Year Yield Spread JPY S&P500
2 Year - 3Month Yield Spread 10 Year Yield US Inflation
Copper / Gold Ratio Gold US Non-farm Payroll
S&P 500/ Dow Jones Ratio Copper US GDP Growth
S&P 500/ Russell 2000 Ratio Oil

Table 2: List of macro-economic indicators used

Figure 1: Walk-forward Optimization: In the optimization process, any point in the future does
not have a bearing on past returns. This reduces the chances of over-fitting significantly. Instead
of finding the best model, walk-forward optimization focuses on finding the best scheme of finding
models that perform well in out of sample data.

3 Model Specification
We use feed-forward neural network architecture with the layer size getting smaller as you go
deeper in the network (dropout was applied to each layer)
• The weight initialization mentioned in [Glorot and Bengio(2010)] was used for each layer
• We used inverted dropout just in case we wanted to allow for changes in dropout rate across
training and testing instances
• To train the model we used mean squared error along with L2-norm regularization on all the
weights as our loss function
• Adam optimizer was used to train the model
• The target variables are future returns over 130 days

4 Training Procedure
Training is done in a walk-forward manner, which basically means that at the time of training the
parameters are learned without having any look-ahead bias. In other words, weights learned on
a certain day are not affected by any data on that day or after that day. Figure 1 depicts how
this works in practice. We have defined following hyper-parameters for the walk-forward training
procedure.
• Initial number of days required for training - This was fixed at 500 days.
• Number of epochs for initial training.

4
 Hyper-parameter Final Values
Expected Returns Model Architecture [Inputs, 72, 48, 32, Outputs]
Dropout probability 0.50
Softmax Layer Architecture [Inputs, 4, Outputs]
# of Monte Carlo simulations 1000
Initial Training Data 500 days
Initial Training Epochs 100
Update Frequency 30
Model Update Epochs 10
Expected Returns Model Prediction Horizon 130 days

Table 3: List of all the hyperparameters with the final values used. For some hyperparameters like
dropout probability or softmax layer architecture we did not search a lot. Wherever possible, we
have used model accuracy to optimize hyperparameters.

• Number of epochs to train the model as it starts walking forward. This controls the extent
to which one wants the network to adapt to new data as we walk forward in time.

• Update frequency - in how many days do we update our model to incorporate new data.
• We use batch-gradient descent over an expanding window as we walk-forward in time. In
other words, every time we retrain the model, we use all the historical data available at that
point in time
• The weights at each training interval are retained as the initial weights for the next training
interval.
We use historical data since 1995. Data till the end of 2013 is used for training and validation
purposes and the remaining as the test data that is only used to evaluate the results once the
investment strategy has been fixed. The hyperparameters are optimized based on the walk-forward
error only and not strategy performance so as to limit overfitting to the strategy performance.
Moreover, strategy performance could potentially be affected by other hyperparameter settings. 1 .
Table 3 shows the list of hyperparameters that are optimized and the final values obtained. An
exponentially decaying learning rate with a minimum learning rate after our decay window expires
is used. Every time we retrain our model we calculate a new mean and standard deviation for our
data to normalize it using a z-score (applied column-wise). We use these same means and standard
deviations to normalize all new data points that come in until the next time we have to retrain our
model, at which time we will calculate a new mean and standard deviation to normalize our data.
At the time of prediction we compute two quantities - expected future returns over the predic-
tion horizon and uncertainties of the prediction. Dropout is a popular regularization technique for
training deep neural networks. Dropout is used in the training stage to improve generalization per-
formance [Srivastava et al.(2014)Srivastava, Hinton, Krizhevsky, Sutskever, and Salakhutdinov]. In
this work, Dropout has also been used as a Bayesian approximation to derive model uncertainty
along with point estimates [Gal and Ghahramani(2016)].
• We use Monte Carlo dropout over M simulations for inference. As it happens while training,
each simulations leads to certain connections in the neural network getting dropped randomly
leading to different estimates. 2

• Expected future return is defined as the mean of the predictions across simulations.
• Expected uncertainty of prediction is defined as the standard deviation of predictions across
simulations. Let’s call this uncertainty measure, dropout uncertainty.
In addition to the dropout uncertainty computed directly from the model, we contrast it with
realized volatility as measure of uncertainty.
1 While intuitively, a model with better accuracy should also lead to better strategy, in practice, this is often not

the case because performance might be dictated by brief periods of high volatility and prediction accuracy is an
aggregate metric
2 M was set to 1000 for all simulations

5
Figure 2: The architecture for the utility function based portfolio construction. The feed-forward
network in blue is used to output expected returns which is then in the first method of portfolio
construction. The uncertainty estimates in green are either obtained from Monte Carlo simulations
using dropout approximation or they are simply obtained from the volatility of respective asset
classes. The purple block of error correction are the prediction errors made by the expected returns
model in blue. Expected returns, uncertainty estimates and errors are fed into a small 1-layer neural
network where the allocations are obtained through a soft-max layer on top.

5 Backtesting
We use daily adjusted prices for backtesting. Both the risk parity benchmarks were implemented
with the same parameters which are highlighted below
• Threshold based rebalancing is with a threshold of 5%.
• Transaction costs of 10bps is used for all backtests unless stated otherwise
• Slippage costs are set at 10% worse than the open price, where 10% is measure in terms of
the difference between daily open and daily close
It is important to note that in utility based portfolio construction backtested results need to be
computed quite often since that is directly used an objective function, therefore while computing
backtested results for optimization purposes we simply assume daily rebalancing since that leads
to faster computation of the objective function.

6 Portfolio Construction
6.1 Based on Expected Returns Model
Once we have expected returns and uncertainty estimates, we use these to tactically change the
weights of an existing allocation within specified limits. The existing allocation could be dollar-

6
based like a market-cap based allocation or a risk based allocation like risk parity. We take risk
parity as the underlying allocation. It is important to note that the following position sizing logic is
agnostic to the underlying allocation and can therefore be used with any static strategic allocation.
• We first divide securities into two buckets - high conviction and low conviction. The securities
are placed in the high conviction bucket if uncertainty is less than a certain threshold and
placed in the low conviction bucket otherwise.
• For securities in the high conviction bucket, we overweight the predictions in the high-
conviction bucket using the following factor, where i is a particular ETF3
 2
µi
Fi = 1 + tanh
σi
• The high conviction weights for each of the underlying assets are equal to their default
underlying allocations multiplied by their respective factor intensity score (Fi )
• The cumulative weight of the high conviction bucket is then equal to the minimum of the
sum of all the high conviction weights and 100 percent
• Once we know the high conviction bucket weight (HW), we can calculate the low conviction
bucket weight as LW = 1 - HW
We add a final layer of error correction to boost performance. We calculate an exponentially
weighted mean absolute error (denoted by δ) for predictions on each of the underlying assets we’re
trading. This allows us to discount predictions that haven’t been correct for a period of time, and
trust predictions that have been right for a period of time. If the errors have been too high recently
(above a threshold), this asset will receive an allocation of 0. As a result, our final allocation within
the high conviction bucket is:
 
µi
σi
Ai = HW ×
δi
The final allocation in the low conviction bucket is the risk-parity allocation for the remaining
assets.

6.2 Based on Utility Function

In this portfolio construction, we add a small dense neural network on top of our mean predictions,
uncertainty, and exponentially weighted mean absolute error with soft-max layer at the top for
portfolio allocations. The architecture is as shown in Figure 2. The soft-max function ensures that
all portfolio allocations sum up to 1. The objective function for training this network is simply the
utility function. The utility function we use in this case is simply the Sharpe Ratio of portfolio
allocations inclusive of all transaction and slippage costs. In order to make sure that the portfolio is
not very concentrated, we add term for entropy across portfolio allocations which favors allocations
whose weights are spread out rather than being concentrated on a few securities.

7 Results
One of the most interesting aspects of our neural network architecture is the Bayesian approxi-
mation. Intuitively, the model uncertainty in expected returns should be a function of expected
volatility, since higher volatility would mean higher inherent uncertainty in expected returns. Ta-
ble 4 shows the correlation of model uncertainties, derived from Monte Carlo simulations using the
dropout at the time of inference, with an estimate of the expected uncertainty. The uncertainty es-
timates seem to be more correlated for volatile asset classes. The low magnitude of the correlation
values suggests that these uncertainty estimates might be independent and a plausible explanation
could be that the model uncertainty captures the uncertainty in modeling future expected returns
as captured in the weights of the neural network whereas the volatility uncertainty captures the
inherent volatility of expected future returns irrespective of the model used.
3 This caps the overweight to a factor of 2x

7
 Security Correlation between DU and VU
BND 0.08
BNDX -0.16
VWOB -0.17
VTI 0.36
VGK 0.30
VWO 0.15
VNQ 0.50
VNQI 0.54

Table 4:
Correlation between uncertainties obtained from the model using dropout approximation and
uncertainty obtained using volatility in different asset classes.

Figure 3: Performance of different portfolio construction methods based on expected returns and
its comparison with risk parity. Volatility Estimate refers to the position sizing where volatility is
used as the uncertainty estimate whereas MC dropout refers to the position sizing where variance
over Monte Carlo simulations is used as volatility estimate. Error Adjusted vs not Error Adjusted
depends on where error correction is applied respectively

Figure 4: Table showing performance metrics of different position sizing heuristics. The position
sizing logic with volatility as the uncertainty estimate and error correction has the best performance
which other heuristics don’t perform as well. Error correction helps reduce maximum drawdown
significantly.

8
Figure 5: The first figure shows mean exponentially weighted MAE (mean absolute error) across
all assets (scaled between 0 and 1). The second figure shows asset allocation plot after error
correction. The third figure show asset allocation using expected returns and volatility based
uncertainty estimate. 2008 financial crisis is particularly interesting - allocation to safe assets
increases. Moreover, model accuracy gets increasingly worse and error correction kicks in and moves
the allocation to cash thereby reducing the maximum drawdown experienced by the portfolio.

9
Figure 6: While error correction helps in reducing drawdown by allocating to cash, the same does
not happen with naive estimates of expected returns. We contrast the performance against an ex-
pected return model where expected return is mean of all historical returns with different thresholds
for moving to cash. It shows that deviation from expected returns is especially informative in case
of our feed-forward neural network model

Figure 7: Performance across different transaction cost assumptions for position sizing using ex-
pected returns model. Although we have worked with conservative slippage assumptions of 10 bps,
we show that the strategy is not very sensitive to the transaction costs - one of the main concerns
for large asset allocators

10
Figure 8: Performance across different slippage assumptions for position sizing based on expected
returns model. Although we have worked with a fairly conservative slippage assumptions of 10%
from open price (10% of the difference between day’s open and day’s close against the side of the
order), the strategy fairly insensitive to these assumptions

Figure 3 shows the performance of different overlays on top of the risk parity strategy. The
volatility based uncertainty estimated with error correction has the best performance. The max
drawdown is significantly lower than risk parity as shown in figure 5, it is mostly due to error
correction. It might appear that the error correction might in itself be causing this gains, but
as shown in figure 6 naive estimates of expected returns don’t show similar benefits with error
correction.
Since we propose an active trading strategy that trades frequently, it is important to take
into account the effect of transaction costs and execution assumptions. Figures 7 and 8 show
performance under different assumptions of transaction costs and slippage.
Figures 9 and 10 show the performance of utility function based portfolio construction using
volatility and dropout variance as uncertainty estimates respectively. The results are in general
better than risk parity. Moreover, it is interesting to note that dropout variance as an uncertainty
estimate does significantly better in terms net returns while having somewhat worse Sharpe Ratio.
Another important observations was the variance in performance across different runs of the neural
network. Although, the aggregate prediction accuracy of the expected returns model did not change
much, the resulting strategy’s performance changed significantly across different runs. 4
Finally, the figure 11 shows the performance in hold out data. It is important to note that while
all the training/validation was done in a walk-forward manner in the period till 2014, this data
was not used at all and all hyperparameters of the training model and strategy were fixed before
running the strategy on this period. One important observation is that in this period, although
there are some significant equity market drops, there is no recession. Major macro indicators such
as Unemployment Rate or yield curve inversion were not particularly useful in detecting the equity
market drops as they were in the period prior to 2014. The strategy with position sizing is based
on the expected returns model with error correction does much worse because of cash allocation
due to error correction. The performance of the utility function based allocations however hold
continue to perform well in this period.

8 Conclusion
Following are some key takeaways of this research
4 We tried to reduce the number of parameters in the model to have lower variance runs, but that resulted in

worse median performance. Moreover it did not help much in terms of reducing the variance

11
Figure 9: Results of the portfolio construction using utility function where we take volatility as
uncertainty estimate. The first figure plots the returns of 5 different runs of the walkforward model
learning procedure against risk parity. The variance comes from different initialization settings.
The second figure shows the asset allocation for one run and the last figure tabulates the results
of the all the runs. The utility based portfolio construction leads to much better performance
compared to risk parity.

12
Figure 10: Results of the portfolio construction using utility function where we take Monte Carlo
dropout variance as the uncertainty estimate. The first figure plots the returns of 5 different runs
of the walkforward model learning procedure against risk parity. The variance comes from different
initialization settings. The second figure tabulates the performance across these runs and the last
figure shows the asset allocation for a sample run. The utility based portfolio construction leads
to much better performance compared to risk parity. Moreover, using MC dropout variance as
uncertainty leads to higher returns, although with lower Sharpe Ratio when compared to volatility
based uncertainty estimate

13
Figure 11: The results for the period post 2014. Note that while all the simulations presented are
done using walk-forward optimization, certain hyper-parameters were fit using the period up till
2014. Performance of the first method of portfolio construction which is based on certain heuristics
has worse performance in this period as compared to others partly because of error correction which
leads to some cash allocation

14
 • We showed a walkforward implementation of a feed-forward neural network that uses macro-
economic indicators and price-volume features to do tactical asset allocation.
• We propose an approach of deriving allocation directly as part of the neural network and
optimizing the weights so as to maximize a custom utility function.
• We show that a deep neural network trained in a walk-forward manner with no look-ahead
bias shows comparable performance to thoroughly researched and hand-crafted strategies
such as risk parity.

9 Future Work
We have identified three main directions for future work.
• Use of hypothetical data : Lack of data makes walk-forward training particularly infeasible.
One of the major issues we saw in the current implementation was significant variance in
performance across different random seed settings. More data would help with convergence.
Data augmentation has had remarkable success in some other settings for deep neural net-
works. We expect either generating simulated data, or using high frequency data as a proxy
should help learning better features for the price-volume data which can then be used in
conjunction with macro-data to get improved performance.
• More macro indicators : While we have used a number macroeconomic data, including more
economic indicators could be useful. Moreover, most economic data releases are typically
more dense than just single numbers . For instance, US Non-farm payroll numbers come out
along with a host of other numbers such as wage growth and participation rate.
• Other measures for uncertainty : Given the low signal to noise ratio nature of financial data
sets, we strongly believe that model derived uncertainty estimates are important. While
in this work model uncertainty did not produce desired results, further work is needed to
ascertain the reason for the same. Alternative ways of addressing model uncertainty like
variational methods also need to be explored.

References
[Bao et al.(2017)Bao, Yue, and Rao] Wei Bao, Jun Yue, and Yulei Rao. A deep learning framework
for financial time series using stacked autoencoders and long-short term memory. PLOS ONE,
12(7):1–24, 07 2017. doi: 10.1371/journal.pone.0180944. URL https://doi.org/10.1371/
journal.pone.0180944.
[Abe and Nakayama(2018)] M. Abe and H. Nakayama. Deep Learning for Forecasting Stock Re-
turns in the Cross-Section. ArXiv e-prints, jan 2018. URL https://arxiv.org/abs/1801.
01777.
[Alberg and Lipton(2017)] J. Alberg and Z. C. Lipton. Improving Factor-Based Quantitative In-
vesting by Forecasting Company Fundamentals. ArXiv e-prints, November 2017.
[Liew and Mayster(2017)] Jim Kyung-Soo Liew and Boris Mayster. Forecasting etfs with machine
learning algorithms. The Journal of Alternative Investments, 20(3):58–78, 2017. ISSN 1520-
3255. doi: 10.3905/jai.2018.20.3.058. URL http://jai.iijournals.com/content/20/3/58.
[Brinson et al.(1995)Brinson, Hood, and Beebower] Gary P. Brinson, L. Randolph Hood, and
Gilbert L. Beebower. Determinants of portfolio performance. Financial Analysts Journal,
51(1):133–138, 1995. doi: 10.2469/faj.v51.n1.1869. URL https://doi.org/10.2469/faj.
v51.n1.1869.
[Blitz and van Vliet(2008)] D.C. Blitz and P. van Vliet. Global Tactical Cross-Asset Allocation:
Applying Value and Momentum Across Asset Classes. ERIM Report Series Research in Man-
agement ERS-2008-033-F&A, Erasmus Research Institute of Management (ERIM), ERIM is
the joint research institute of the Rotterdam School of Management, Erasmus University and
the Erasmus School of Economics (ESE) at Erasmus University Rotterdam, June 2008. URL
https://ideas.repec.org/p/ems/eureri/12598.html.

15
[Kelly(1956)] J. L. Kelly. A new interpretation of information rate. The Bell System Technical
Journal, 35(4):917–926, July 1956. ISSN 0005-8580. doi: 10.1002/j.1538-7305.1956.tb03809.x.
URL https://www.princeton.edu/~wbialek/rome/refs/kelly_56.pdf.
[Breiman(1961)] L. Breiman. Optimal gambling systems for favorable games. In Proceedings
of the Fourth Berkeley Symposium on Mathematical Statistics and Probability, Volume 1:
Contributions to the Theory of Statistics, pages 65–78, Berkeley, Calif., 1961. University of
California Press. URL https://projecteuclid.org/euclid.bsmsp/1200512159.

[Chakravorty(2017)] Gaurav Chakravorty. Choosing between mean variance, kelly and risk parity
in asset allocation. Qplum Technical Articles, December 2017. URL https://www.qplum.co/
documents/december-2017-market-commentary.
[Louton et al.(2015)Louton, McCarthy, Rush, Saraoglu, and Sosa] David Louton, Joseph Mc-
Carthy, Stephen Rush, Hakan Saraoglu, and Ognjen Sosa. Tactical asset allocation for
us pension investors: How tactical should the plan be? Journal of Asset Management,
16(7):427–436, Dec 2015. ISSN 1479-179X. doi: 10.1057/jam.2015.26. URL https:
//doi.org/10.1057/jam.2015.26.
[Singhal and Chakravorty(2018)] Mansi Singhal and Gaurav Chakravorty. Investment objective
is not only about maximizing returns. Qplum Technical Articles, Jan 2018. URL https:
//www.qplum.co/documents/january-2018-market-commentary.
[Estrella and Hardouvelis(1991)] Arturo Estrella and Gikas A. Hardouvelis. The term structure
as a predictor of real economic activity. The Journal of Finance, 46(2):555–576, 1991. ISSN
00221082, 15406261. URL http://www.jstor.org/stable/2328836.

[Estrella and Mishkin(1996)] Arturo Estrella and Frederic S. Mishkin. The yield curve as a
predictor of u.s. recessions. New York Federal Reserve Research Articles, 1996. URL
https://www.newyorkfed.org/research/current_issues/ci2-7.html.
[Kliesen(2018)] Kevin L. Kliesen. Recession signals: The yield curve vs. unemploy-
ment rate troughs. Federal Reserve Bank of St Louis Research Articles, June
2018. URL https://research.stlouisfed.org/publications/economic-synopses/2018/
06/01/recession-signals-the-yield-curve-vs-unemployment-rate-troughs/.
[Smalter Hall(2017)] Cook, Thomas-R Smalter Hall, Aaron. Macroeconomic Indicator Forecasting
with Deep Neural Networks. Federal Reserve Bank of Kansas City Working Paper, September
2017.

[Glorot and Bengio(2010)] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of train-
ing deep feedforward neural networks. Proceedings of the Thirteenth International Conference
on Artificial Intelligence and Statistics, 2010. URL http://proceedings.mlr.press/v9/
glorot10a/glorot10a.pdf.
[Srivastava et al.(2014)Srivastava, Hinton, Krizhevsky, Sutskever, and Salakhutdinov] Nitish Sri-
vastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov.
Dropout: A Simple Way to Prevent Neural Networks from Overfitting. J. Mach. Learn. Res.,
15(1):1929–1958, January 2014. ISSN 1532-4435. URL http://portal.acm.org/citation.
cfm?id=2670313.
[Gal and Ghahramani(2016)] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approxi-
mation: Representing model uncertainty in deep learning. In Proceedings of the 33rd Interna-
tional Conference on International Conference on Machine Learning - Volume 48, ICML’16,
pages 1050–1059. JMLR.org, 2016. URL http://dl.acm.org/citation.cfm?id=3045390.
3045502.

16
 Disclosures
All investments carry risk. This material is intended only for institutional investors, investment professionals, market
counterparts or intermediate customers and may not be reproduced or otherwise disseminated in whole or in part without
prior written consent.

This document has been provided to you solely for information purposes and does not constitute an offer or solicitation
of an offer or any advice or recommendation to purchase any securities or other financial instruments and may not be
construed as such. It is not an offer or a solicitation for the sale of a security nor shall there be any sale of a security in
any jurisdiction where such offer, solicitation or sale would be unlawful. The factual information set forth herein has been
obtained or derived from sources believed to be reliable but it is not necessarily all-inclusive and is not guaranteed as to
its accuracy and is not to be regarded as a representation or warranty, express or implied, as to the information, accuracy
or completeness, nor should the attached information serve as the basis of any investment decision. Past performance is
not indicative of future performance.

This presentation may contain hypothetical performance results. Hypothetical performance results, while imminently useful
in understanding the merit of the methodology, have many inherent limitations, some of which are described below. No
representation is being made that any account will or is likely to achieve profits or losses similar to those shown. In
fact, there are frequently sharp differences between hypothetical performance results and the actual results subsequently
achieved by any particular trading program.

One of the limitations of hypothetical performance results is that they are generally prepared with the benefit of hindsight.
In addition, hypothetical trading does not involve financial risk, and no hypothetical trading record can completely account
for the impact of financial risk in actual trading. For example, the ability to withstand losses or adhere to a particular
trading program in spite of trading losses are material points which can also adversely affect actual trading results. There
are numerous other factors related to the markets in general or to the implementation of any specific trading program
which cannot be fully accounted for in the preparation of hypothetical performance results and all of which can adversely
affect actual trading results.

Investing in futures, derivatives or foreign exchange markets is highly speculative and involves substantial investment,
liquidity and other risks. CTA managed accounts and hedge funds can be leveraged and their performance results can be
volatile. Past performance of issuers, financial instruments and markets may not be indicative of future results, and there
is no guarantee that targeted performance will be achieved.

Please visit our website for full disclaimer and terms of use.

17

View publication stats