Deep sector rotation swing trading

Joel R. Bock1 and Akhilesh Maewal2

1
Independent, New Braunfels, Texas, USA
2
Independent, San Diego, California, USA

January 4, 2023

Abstract
A system for sector rotation swing trading of exchange-traded funds (ETFs) using deep learning is presented.
Weekly trades are made on funds representing 11 major sectors of the U.S. economy. The trading system
was backtested for the period January 2012 through December 2022. Annualized CAGR returns exceeded the
benchmark buy-and-hold strategy by an average 12.63% (median 7.63%). Of particular note is the positive alpha
(α = 28.4%) achieved in trading for 2022, a difficult year for stocks in which the S&P 500 index experienced a
CAGR loss of 18% .Over the studied period, Sharpe ratios averaged 1.39, and the mean maximum drawdown
was 10%. The deep model design is multiple-input, multiple output, and can be easily extended to include other
factors that may influence predictability of future price movements. The results presented here are preliminary,
and are exclusive of trading costs. Analysis of these costs is prerequisite to deployment as a semi-mechanical
swing trading system.

1 Introduction
Sector rotation refers to a portfolio rebalancing strategy wherein money is moved from one economic sector into
another, to profit from differential financial performance of different sectors mapped to stages of the business cycle
[3]. The cycles follow the sequence: [trough, expansion, peak, contraction], before repeating again. The idea is to
identify and anticipate the timing of each phase based on macroeconomic indicators, and allocate assets accordingly
[10].
For quantitative trading, business cycle timing is problematic. Incorrect prognosis may result in trading out-
of-phase, yielding capital losses, or at least underperformance relative to a simple long position in a broad market
index. In studies of quarterly switching between index funds of two asset classes (U.S. stocks and bonds) over the
period 1993-2017, it was concluded that the optimal timing of rotation between classes was ``indistinguishable from
a random sequence" [16]. Even given perfect forecasting of cycles, researchers found at best a 2.3% outperformance
relative to the market using a conventional sector rotation strategy [12]. Their study covered multiple economic
expansions and contractions over the years 1948–2007.
Industrial momentum strategies for sector rotation eliminate the need to time the market, and may produce
outperformance, especially when rebalanced in the short term [17]. In one recent study, industrial momentum
strategies with various holding periods were examined on a large historical sample spanning years 1926-2018. U.S.
industry portfolios were sorted into quintiles of returns from the previous month. Portfolios comprising industries
with the highest previous‐month returns were found to produce significantly greater profits than the groups having
underperformed in previous month [11].
The present work takes a different strategic approach to short-term sector rotation trading. Deep learning
models are developed to inform buy decisions on exchange-traded funds (ETFs) representing the major sectors U.S.
economy. For each trading year, a multi-input, multi-output model is trained on recent historical price, volume and
auxiliary economic data to make recommendations on which sectors are most likely to increase in value over the
next trading week. Selected funds are ranked, curated and bought at Monday market open, and all holdings are
liquidated at close of the current trading week. Within-year models are incrementally updated each week. In the
short term, ETF price movements exhibit volatility; this can be exploited for tactical asset allocation.
Results of backtesting the current trading strategy are presented for the period January 2012 through December,
2022. Exclusive of trading costs, CAGR outperformance over the benchmark S&P 500 index was found to be
meaningful (mean α = 12.63%, median = 7.63%).

1
Electronic copy available at: https://ssrn.com/abstract=4280640
Related work
Machine and deep learning techniques as applied to portfolio allocation and trading are abundant in the literature.
A selected review of reports relevant to the present work is provided here.
Navon and Keller [18] developed a deep learning-based system to predict price trends of stocks and ETFs. These
predictions were used for intraday trading using a ``buy-sell-hold" strategy. Potential open positions were evaluated
for likelihood of profitability by thresholding probabilistic neural network outputs. Positions were closed based on
forecast changes in price trend. Backtesting over two years, cumulative returns from trading certain large-cap stocks
generally outperformed broad market index. The present research was informed by [18] where it was suggested that
prediction accuracy could be improved by making dynamic model updates using all available at a point in time.
This idea is implemented in the strategy reported here.
In Liew and Mayster [15], ETF price movements were forecast using machine and deep learning models. Predic-
tion horizons of one and three months' duration were found to produce best results in terms of a ``gain criterion"
comparing model price forecast accuracy against random inputs. No explicit financial results were reported. One
conclusion was that ETF price movements were essentially unpredictable at very short horizons (one to five days)
[15]. Findings in the present report contradict this assertion; it will be shown that over a one-week trading horizon,
profitability is indeed possible using a deep learning strategy.
Chalkidis and Savani [5] applied several machine learning techniques (including random forests and deep net-
works) to financial time series prediction. Models were trained to output one of three decisions on price movement--
down, up, or slightly in either direction, corresponding to short, long and abstain positions. A trading strategy for
commodities futures based on these ``selective classification" models with different feature sets was designed, and
subjected to backtesting. By including an ``abstain" position, the selective classification approach outperformed
non-selective (binary) classifiers, and resulted in smaller capital losses in trading due to reduced misclassifications.
Pinelis and Ruppert [20] used random forest models for portfolio adjustment on a monthly basis. Their models
simultaneously forecast both expected returns and volatility. Results produced improved returns (α = 3.37%) over
a buy-and-hold strategy. The authors speculate that further performance gains might be achieved by use of deep
learning.
Sauer [21] studied investment strategies, using random forests to identify economic regimes based on monthly
GDP data. Portfolio sector weights between cyclical and defensive assets were modified according to the identified
regime. It was concluded that for trading, machine learning did not significantly improve results compared to a
naïve strategy using equal weights for sectors.
Additional work on deep learning methods for financial applications is surveyed in [19].

Present contribution
The key contributions made in this research are proposed as follows:

• A robust system for sector rotation swing trading using deep learning. A single model architecture, set of
training hyper-parameters and trading rules generalizes to all years considered (2012-2022).
• Consistent outperformance over the benchmark buy-and-hold strategy.
• A statistically-based estimate of confidence in each putative trade.

• Coverage includes all major sectors of U.S. economy, with each sector potentially bought in a trading week.
Custom models are developed within each year, and updated dynamically using the latest observed price and
volume data.
• The deep model design is multiple-input, multiple output and extendable to include other technical indicators
or economic factors that may impact future price movements.

2 Methods
The trading system comprises three elements: (1) historical price and volume data for major economic sector ETFs;
(2) a deep learning model trained to forecast future price movements using this data; and (3) a trading strategy
based on the out-of-sample predictions made by the model.

2
Electronic copy available at: https://ssrn.com/abstract=4280640
2.1 Data preparation
The ETFs selected as the basis for development of the trading strategy in this study are listed in Table 1. As a
whole, these funds cover the major industrial sectors of the United States economy. Price and volume data for each
fund were acquired from Yahoo Finance† .
In addition to the ETF price and volume data, ancillary economic data (10 year U.S. Treasury yield, USD
currency index, crude oil proxy and market volatility indicators) were collected, under the hypothesis they might
provide additional informative context to the model. In some experiments, these extra data were as model inputs,
and were not actively traded quantities. The final results reported here are based only on model inputs comprising
the sector funds of Table 1.
Sector ETF and auxiliary data were sampled at discrete points in time, and filtered to exclude all but ``Friday
close" prices. Non-trading days were removed from the data.
Let the cardinality of traded ETFs and auxiliary input data be represented by l and m respectively. Training
examples for week t are constructed by assembling matrices containing prices, volumes and other data recorded for
the current and previous N − 1 weeks. Each input data matrix Xt has dimensions N × (l + m)‡ . Length l binary
label vectors yt+1 describing the percentage increase in prices in week t + 1 are assigned to each matrix, resulting
in a set of examples z = (Xt , yt+1 ).
All non-target data were normalized to have zero mean and unit variance.
The present swing trading scheme liquidates all holdings upon weekly market close. Training labels y signify
price movement for the succeeding week-- an uptick in price signals that the associated ETF should be bought.
A threshold increase of 100 basis points was used to assign these labels for the classification task posed to the
model during training. By defining a target value in this manner, the trading model was charged with learning to
implicitly predict two distinct future values (Monday open and Friday close).
Input data are shown schematically in Figure 1. For clarity, quantities not subject to forecast by the trained
model (volumes and optional accessory input data) are not represented in this figure.

Sector name Symbol

Information Technology XLK
Health Care XLV
Consumer Discretionary XLY
Communication Services VOX
Financials XLF
Industrials XLI
Consumer Staples XLP
Utilities XLU
Materials XLB
Real Estate IYR
Energy XLE

Table 1: Sectors and their ETF symbols.

2.2 Deep learning model

A deep learning network model was developed to forecast future price movements of ETFs and inform swing trading
decisions. The model was designed to process multiple inputs and make predictions for each sector-based asset as
described in Section 2.1. After considerable experimentation with different architectures, the final model used here
was composed of four fully connected internal layers, each followed by ReLU activation [14] and dropout [24], now
standard techniques used to mitigate vanishing gradients and over-fitting on the training data, respectively. The
output layer was likewise densely connected, and was terminated with a linear activation function.

Model optimization and dynamic update

Individual models were trained and evaluated for trading years from January 2012 through December 2022 . Each
model was fit using historical data from the previous 2 years. After training, the models were used to recommend
purchases of sector ETFs once per week in the succeeding trading year. All current holdings were liquidated just
† https://finance.yahoo.com
‡ The dimensions are N × (2l + m) if ETF volumes are used in the analysis.

3
Electronic copy available at: https://ssrn.com/abstract=4280640
Figure 1: Input data scheme for the trading model. Columns represent traded assets as listed in Table 1. Each box
along a row represents a Friday closing price xt for week t. N is the history depth. In this example, the deep model
will learn to suggest trades to be executed during the following week t + 1. A complete data example comprises
one such frame of data and a target label vector containing one number per traded asset. Volumes and optional
accessory data (non-forecast quantities) are omitted for clarity.

prior to closing of the trading week, and profits & losses recorded. Once trades were completed, the models were
updated incrementally using the observed price data from the week just ended, and used to generate signals for
ETF buys in the following week.
To connect dynamic model optimization with financial performance, a custom loss function incorporating recent
capital gains and losses was used to steer model weights updates in the direction of profitable trades. The idea of
directly using a financial criterion was suggested in [4].
Models were developed used the TensorFlow platform [1] and the Keras deep learning API [6].

Evaluation of trading performance

At trading year-end, overall performance of executed trades was evaluated using financial performance statistics
including compound annual growth rate (CAGR), the Sharpe ratio [22], and the maximum drawdown of the
portfolio. Present results are compared against the benchmark strategy of full investment in the S&P 500 (SPX)
for a given trading year. As defined here, the excess return of the deep models DL is

α = CAGRDL − CAGRSP X

where CAGRSP X includes reinvestment of dividends§ . This provided an objective means to assess practical value
added of the machine trading approach.
The risk-free return used to estimate Sharpe ratios was based on contemporaneous 90-day U.S. Treasury yields.
All trading was performed on out-of-sample data.

2.3 Swing trading

The strategy for weekly sector rotation trading is summarized as follows:
• Immediately prior to Friday closing (week t):
– Sell: Exit all positions from week t − 1.

• After closing (week t):

– Analyze:
1. Execute deep model, and select potential ETFs to buy in the next week.
§ Source: https://bit.ly/3iavw0u

4
Electronic copy available at: https://ssrn.com/abstract=4280640
 2. Assign confidence metric to distribution of predictions for each asset.
3. Apply loss reduction heuristics to refine the selection set.
4. Rank funds in list, and allocate available capital.
5. Update model using the observed week data xt .
• On Monday open (week t + 1):

– Buy: Buy sector ETFs as designated in week t.

Generation of ``buy" signal

The trading model produces a vector of real numbers, one for each sector ETF from the collection listed in Table 1.
A preliminary buy decision is made by comparing each output value to a threshold value unique to the corresponding
asset. Surpassing this value is taken as a ``buy" signal; otherwise, the fund is not bought. The threshold vector
is updated dynamically after each trading week, and is estimated by optimizing a cost function based on receiver
operating characteristic (ROC) curve analysis [8] of recently observed input data.

Scheduled selling
All positions are liquidated just before market close in the current trading week. While perhaps counter-intuitive,
this systematic approach eliminates potential situations where even the professional investor's selling decisions may
perform worse than a random selling strategy [2].

Confidence estimation
It has been noted that neural network predictions can be ``…very noisy and unreliable"[4]. This assertion was
corroborated in the present research, where point estimates of future price movements were observed to be highly
variable, and distributed non-uniformly over successive experimental runs. A confidence measure was implemented
by using a technique termed ``Monte Carlo dropout" [9] applied to the output layer in the model. A large number of
predictions is made for each asset, for each trading week; each prediction is made with a random group of connection
weights nullified. In effect, this produces an ensemble of deep models. For each asset s, the statistical distribution
of predictions ŷs for the ensemble is analyzed. Greater confidence is assigned to forecasts where a significant portion
(here, 80%) of the probability mass of the distribution P̂s is located within one standard deviation of the population
median. A dispersed distribution in contrast presents greater uncertainty, and the associated asset is not purchased.

Loss reduction heuristics

Greater psychological discomfort accrues from losses than does pleasure realized from equivalent financial gains [25].
Loss aversion strategies are essential for swing trading to limit short-term losses. Additional filters may be applied
to the ETFs marked for purchase in the ensuing week. Filters are parameterized to align with trader's level of loss
aversion. These include loss limits at the single trade or complete portfolio level. Actions taken upon exceedance
may dictate removal of a symbol from the buy list, or halting of all trading for the current week.
Stop limiting conditions and consequent strategic actions used in the backtesting experiments are summarized
in Table 2.

Condition Value Time Action

Recent loss by symbol 5% week Remove ETF from buy list
Maximum loss, week-week $300 week Halt trading one week
Portfolio underwater 5% Q4 Halt trading one week
Maximum loss by symbol 27.5% Q1-Q4 Remove ETF from buy list
Minimum win rate by symbol 45% Q4 Remove ETF from buy list

Table 2: Loss mitigation conditions and resulting actions used in backtesting experiments.

Dynamic allocation of capital

Prior to a trading week, available capital is distributed to each sector fund based on total average investment
performance to date (win rate), and on momentum of trades in the sector [17]. Define a ``win" as a profitable
ETF trade at close of the position. Allocation weights for sector s in the next period are computed by the equation

5
Electronic copy available at: https://ssrn.com/abstract=4280640
ws = 1.0 + wins streaks
buyss + winss +1 where winss and buyss are the total win count and number of weeks fund s was bought,
s

respectively; streaks is the number of consecutive buys producing a profitable outcome (not necessarily contiguous
weeks).

3 Results and discussion

Trading performance
Main results from backtesting the current trading strategy are presented in Table 3. Deep sector rotation produced
on average excess returns α of 12.63% per year (median = 7.63%) over the benchmark for 11years studied. The
strategy produced positive excess returns each year, the singular exception being 2012, where underperformance
was 75 basis points. Of particular note is the positive alpha (α = 28.4%) achieved in trading for 2022, a difficult
year for stocks in which the S&P 500 index experienced a CAGR loss of 18% .
Sharpe ratios averaged 1.39, indicating reasonable return-to-risk from trades execution. Overall, ∼ 60% of
executed trades were found profitable. Maximum drawdown is observed to be, on average, roughly 10%. The
right-most column in Table 3 displays selectivity of the trading system--only 2.25 out of 11 potential sectors are
bought per week. This is a result of the procedures used for filtering putative trades as discussed in Section 2.3.
Specific sectors traded for each year in the backtesting period are summarized in Appendix A.

Year CAGR α Sharpe MaxDD Wins #Buy/wk.

Current SPX
2012 15.13 15.88 -0.75 1.76 -0.11 60.51 3.02
2013 47.57 32.43 15.14 1.87 -0.09 63.91 3.31
2014 21.44 13.81 7.63 1.09 -0.07 61.90 1.65
2015 2.72 1.31 1.41 1.46 -0.09 50.00 2.78
2016 54.41 11.93 42.48 1.36 -0.07 63.11 1.98
2017 22.48 21.94 0.54 1.61 -0.06 63.08 2.50
2018 -3.18 -4.41 1.23 0.94 -0.15 54.55 1.27
2019 42.46 31.74 10.72 1.91 -0.05 69.33 2.94
2020 25.00 18.38 6.62 0.95 -0.10 61.25 1.57
2021 54.38 28.83 25.55 1.38 -0.07 60.66 2.35
2022 10.29 -18.11 28.40 0.97 -0.21 57.58 1.53
Mean 26.61 13.97 12.63 1.39 -0.097 60.62 2.25
Median 22.48 15.88 7.63

Table 3: Backtesting results for the analyzed period January 2012 through December 2022. CAGR: compound
annual growth rate (%); α: excess return of Current vs. SPX naïve strategies (%); Sharpe: Sharpe ratio; MaxDD:
maximum drawdown; Wins: trade win %; #Buy/wk: average number of ETF trades per week.

Discussion
Trading results from the present deep strategy demonstrate non-trivial outperformance for over one decade of
backtesting. By using only two years' training data, and incrementally refreshing the models with newly observed
data at the end of each trading week, the variance between backtesting and out-of-sample performance can be
mitigated. Previous authors [26] suggested that too many backtests increase overfitting, reducing the generalization
ability of financial forecasts. That observation was based on a ``fit-once, forecast all" approach, which does not
apply to the continuous updating trading scheme followed here.
The densely-connected multiple-input, multiple-output model learns to approximate a function connecting prices
and volumes in each major sector of the economy with their future values. Nonlinear interactions between prices
across (perhaps linearly uncorrelated) sectors are represented. Results obtained here agree with previous findings
of evidence for a universal structure of price formation in high-frequency trading, where deep models trained on all
stocks were seen to outperform stock-specific ones [23].
Regarding labeling of the training data, declaring a minimum 1% increase in next week's price as the target
positive class was arrived at in order to approximately balance the examples for classification. A more accurate
threshold value could be determined with further analysis. The model learns to approximate the target function
even with misclassification errors, in part due to the error signal provided by the loss function which includes capital
gains or losses from the previous weeks' trading.

6
Electronic copy available at: https://ssrn.com/abstract=4280640
 Returns reported here are exclusive of ETF trading costs. Commissions (minimal), changes in underlying net
asset values, operating expenses, and bid/ask spreads all affect the total cost of ETF ownership. For active trading
as considered here, bid/ask spreads are probably the most consequential [7]. Tax liabilities may be incurred from
capital gains in taxable trading accounts. Future extensions of this work should a realistic assessment of transaction
costs, e.g., following analysis reported in Jensen et al. [13].

4 Conclusion
This paper presents a strategy for weekly sector rotation swing trading ETFs using deep learning. Coverage of
potential trades each week includes 11 major sectors of the U.S. economy. A single model architecture, set of
training hyper-parameters and trading rules was used in backtesting covering the period January 2012 through
December 2022. The model design is multiple-input, multiple output and can be easily extended to include other
technical indicators or economic factors that may influence predictability of future price movements.
Based on annualized CAGR, returns exceeded the benchmark strategy on average by 12.63% (median 7.63%).
Sharpe ratios averaged 1.39, indicating reasonable return-to-risk from trades execution. Mean maximum drawdown
of the dynamic portfolio was 10% for the studied period.
The results presented here are preliminary, and not inclusive of real-world trading costs. Analysis of these costs
is obviously prerequisite to deployment as a semi-mechanical swing trading system.
Additional extensions to this work might include more exhaustive search for an optimal set of parameter values
defining the trading model architecture, or variations in training hyper-parameters and loss reduction heuristics
such as those listed in Table 2.

A Appendix
Summaries of trades made by the trading model for years 2012 through 2022. In each table, #Buy and Wins are
counts and win percentages for corresponding ETF trades, respectively.

Symbol #Buy Wins Symbol #Buy Wins Symbol #Buy Wins

XLV 28 64.29 XLF 36 61.11 XLV 17 76.47
XLU 22 77.27 XLK 34 64.71 XLU 15 73.33
IYR 21 52.38 XLB 18 61.11 XLK 13 61.54
XLF 18 66.67 XLU 18 72.22 XLB 12 66.67
XLY 17 64.71 XLE 14 64.29 XLI 11 54.55
XLB 16 56.25 XLI 12 58.33 VOX 5 40.0
XLK 10 60.0 XLY 12 83.33 XLF 3 33.33
XLI 9 44.44 XLP 8 100.0 XLP 3 33.33
VOX 8 50.0 IYR 7 28.57 XLY 3 33.33
XLP 5 40.0 XLV 7 28.57 IYR 1 100.0
XLE 3 33.33 VOX 3 66.67 XLE 1 0.0

Table 4: Trades made for 2012. Table 5: Trades made for 2013. Table 6: Trades made for 2014.

where DL
Symbol #Buy Wins Symbol #Buy Wins Symbol #Buy Wins

XLV 24 54.17 XLU 26 57.69 XLV 21 61.9

XLI 16 56.25 XLF 18 77.78 VOX 16 37.5
XLY 16 43.75 XLV 10 40.0 XLP 16 87.5
XLB 15 33.33 XLY 10 40.0 XLB 12 66.67
XLF 14 42.86 XLE 9 88.89 XLI 12 75.0
XLK 12 66.67 IYR 8 62.5 XLK 12 66.67
XLP 12 58.33 VOX 7 71.43 IYR 11 63.64
VOX 10 60.0 XLP 5 80.0 XLF 11 54.55
XLE 9 44.44 XLB 4 100.0 XLU 10 80.0
XLU 9 44.44 XLI 3 33.33 XLY 6 50.0
IYR 5 40.0 XLK 3 33.33 XLE 3 33.33

Table 7: Trades made for 2015. Table 8: Trades made for 2016. Table 9: Trades made for 2017.

7
Electronic copy available at: https://ssrn.com/abstract=4280640
 Symbol #Buy Wins Symbol #Buy Wins Symbol #Buy Wins
XLV 11 54.55 VOX 23 73.91 XLF 14 42.86
XLF 10 50.0 XLP 22 77.27 XLY 13 84.62
XLU 7 71.43 XLU 21 66.67 XLI 10 80.0
XLE 6 50.0 IYR 19 78.95 XLV 10 40.0
XLP 6 33.33 XLY 17 70.59 IYR 9 77.78
VOX 5 40.0 XLB 12 33.33 XLK 7 71.43
XLI 5 60.0 XLK 9 100.0 XLP 5 40.0
XLK 5 100.0 XLV 9 44.44 VOX 3 66.67
IYR 4 50.0 XLI 8 62.5 XLB 3 66.67
XLB 4 25.0 XLE 6 83.33 XLE 3 33.33
XLY 3 66.67 XLF 4 50.0 XLU 3 33.33

Table 10: Trades made for 2018. Table 11: Trades made for 2019. Table 12: Trades made for 2020.

Symbol #Buy Wins Symbol #Buy Wins

XLE 21 61.9 XLU 15 66.67
XLB 20 55.0 XLE 12 58.33
XLK 18 66.67 XLV 9 66.67
XLI 14 71.43 XLI 7 57.14
XLU 11 27.27 IYR 6 83.33
XLY 9 66.67 XLF 6 50.0
IYR 8 100.0 XLK 5 20.0
XLF 8 37.5 XLP 4 75.0
VOX 6 50.0 VOX 3 66.67
XLV 4 75.0 XLY 2 50.0
XLP 3 100.0 XLB 1 0.00

Table 13: Trades made for 2021. Table 14: Trades made for 2022.

References
[1] M. Abadi et al. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. Software available
from tensorflow.org.

[2] K. Akepanidtaworn, R. Di Mascio, A. Imas, and L. Schmidt. Selling Fast and Buying Slow: Heuristics and
Trading Performance of Institutional Investors. SSRN Electronic Journal, pages 1--92, Sept. 2019. https:
//ssrn.com/abstract=3301277.
[3] A. Beber, M. W. Brandt, and K. A. Kavajecz. What does equity sector orderflow tell us about the economy?
Working Paper 16534, National Bureau of Economic Research, November 2010.

[4] Y. Bengio. Using a financial training criterion rather than a prediction criterion. International Journal of
Neural Systems, 8 4:433--43, 1997.
[5] N. Chalkidis and R. Savani. Trading via selective classification. In Proceedings of the Second ACM International
Conference on AI in Finance, number 14, pages 1--9. ACM, November 2021.

[6] F. Chollet et al. Keras. https://github.com/fchollet/keras, 2015.

[7] E. Doak. ETFs: How much do they really cost? https://bit.ly/3fIRpm6, September 9 2021.
[8] T. Fawcett. An introduction to ROC analysis. Pattern Recognition Letters, 27(8):861--874, 2006.
[9] Y. Gal and Z. Ghahramani. Dropout as a Bayesian approximation: Representing model uncertainty in deep
learning. In M. F. Balcan and K. Q. Weinberger, editors, Proceedings of The 33rd International Conference
on Machine Learning, volume 48 of Proceedings of Machine Learning Research, pages 1050--1059, New York,
New York, USA, 20--22 Jun 2016. PMLR.
[10] T. Greetham and M. Hartnett. The investment clock. Technical report, Merrill Lynch Global Securities
Research & Economics Group, 10 November 2004.

[11] K. Grobys and J. Kolari. On industry momentum strategies. Journal of Financial Research, 43(1):95--119,
2020.
[12] B. Jacobsen, J. S. Stangl, and N. Visaltanachoti. Sector Rotation Across the Business Cycle. SSRN Electronic
Journal, pages 1--53, Dec. 2009. https://ssrn.com/abstract=1467457.

8
Electronic copy available at: https://ssrn.com/abstract=4280640
[13] T. I. Jensen, B. T. Kelly, S. Malamud, and L. H. Pedersen. Machine Learning and the Implementable Efficient
Frontier. SSRN Electronic Journal, pages 1--70, Aug. 2022. https://ssrn.com/abstract=4187217.

[14] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks.
Communications of the ACM, 60:84 -- 90, 2012.
[15] J. K.-S. Liew and B. Mayster. Forecasting ETFs with Machine Learning Algorithms. SSRN Electronic Journal,
Jan. 2017. https://ssrn.com/abstract=2899520.
[16] G. Metcalfe. The mathematics of market timing. PLOS ONE, 13(7):1--12, 07 2018.

[17] T. J. Moskowitz and M. Grinblatt. Do industries explain momentum? The Journal of Finance, 54(4):1249--
1290, August 1999.
[18] A. Navon and Y. Keller. Financial time series prediction using deep learning. Available at arXiv: https:
//arxiv.org/abs/1711.04174, November 2017.

[19] A. M. Ozbayoglu, M. U. Gudelek, and O. B. Sezer. Deep learning for financial applications : A survey. Applied
Soft Computing, 93, 2020.
[20] M. Pinelis and D. Ruppert. Machine Learning Portfolio Allocation. SSRN Electronic Journal, pages 1--42,
Mar. 2020. https://ssrn.com/abstract=3546294.

[21] M. Sauer. Sector rotation through the business cycle: A machine learning regime approach. pages 1--36,
September 30 2019. https://ssrn.com/abstract=3473907.
[22] W. F. Sharpe. Mutual fund performance. Journal of Business, 39(1):119--138, 1966.
[23] J. Sirignano and R. Cont. Universal Features of Price Formation in Financial Markets: Perspectives From
Deep Learning. SSRN Electronic Journal, pages 1--20, March 2018. https://ssrn.com/abstract=3141294.

[24] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov. Dropout: A simple way to
prevent neural networks from overfitting. Journal of Machine Learning Research, 15(56):1929--1958, 2014.
[25] A. Tversky and D. Kahneman. Prospect theory: An analysis of decision under risk. Econometrica, 47(2):263-
-292, 09 1979.

[26] T. Wiecki, A. Campbell, J. Lent, and J. Stauth. All that glitters is not gold: Comparing backtest and out-
of-sample performance on a large cohort of trading algorithms. The Journal of Investing, 25(3):69--80, Fall
2016.

9
Electronic copy available at: https://ssrn.com/abstract=4280640