A Robust Transferable Deep Learning Framework for Cross-sectional Investment Strategy
0 likes396 views
The document discusses the AAAI-20 workshop on knowledge discovery from unstructured data in financial services, presenting a deep learning framework called ric-nn for cross-sectional investment strategies. It outlines the methodology, data used, and experimental results, comparing the performance of ric-nn with other models in stock prediction. The findings highlight the advantages of a nonlinear multi-factor approach, rank IC-based stopping, and multi-region transfer learning for portfolio performance.
![✓ 20 factors:
Problem Formulation
・ Given a stock 𝑖 at month 𝑇 (𝑖 ∈ 𝑈 𝑇: a regional index constituents at 𝑇)
𝒙𝑖,𝑇 ∈ 𝑅20
✓ Features: 20 factors and preprocessed factors
𝑥/ 𝑅
𝑦 ≔ 2(𝑥 − 𝑦)/( 𝑥 + 𝑦 )
✓ Output variable: scaled one-month-ahead stock return
Scale to the range [0,1]
Pre-processing ＆ Feature augmentation
𝒗𝑖,𝑇 = (𝒙𝑖,𝑇, 𝒙𝑖,𝑇−3, … , 𝒙𝑖,𝑇−12, 𝒙𝑖,𝑇/ 𝑅
𝒙𝑖,𝑇−3, … , 𝒙𝑖,𝑇/ 𝑅
𝒙𝑖,𝑇−12) ∈ [0,1] 𝟏𝟖𝟎
𝑟𝑖,𝑇+1 ∈ [0,1]
12
・Most of factors are updated quarterly ・Time difference between the present and each quarter ago](https://image.slidesharecdn.com/kdf2020nakagawa-200205041410/85/A-Robust-Transferable-Deep-Learning-Framework-for-Cross-sectional-Investment-Strategy-13-320.jpg)
![Performance Measure
・ We use the following (standard) measures.
✓ Alpha Return ≔ ς 𝑡=1
𝑇
1 + 𝛼 𝑡
12/𝑇 − 1
✓ 𝑇𝐸 Risk ≔
12
𝑇−1
𝛼 𝑡 − 𝜇 𝛼
2
✓ 𝐼𝑅 Return/Risk ≔ Alpha/𝑇𝐸
portfolio return – benchmark return
𝜇 𝛼: Average of Alpha
𝛼 𝑡:
✓ MaxDD Worst − case Loss ≔ min
𝑘∈[1,𝑇]
(0,
𝑊𝑘
𝑃𝑜𝑟𝑡
max
𝑗∈ 1,𝑘
𝑊𝑗
𝑃𝑜𝑟𝑡 − 1)
𝑊𝑘
𝑃𝑜𝑟𝑡
: Cumulative return of the portfolio
17](https://image.slidesharecdn.com/kdf2020nakagawa-200205041410/85/A-Robust-Transferable-Deep-Learning-Framework-for-Cross-sectional-Investment-Strategy-18-320.jpg)
A Robust Transferable Deep Learning Framework for Cross-sectional Investment Strategy
- 1. The views expressed here are our own and do not necessarily reflect the views of Nomura Asset management. Any errors and inadequacies are our own. The AAAI-20 Workshop on Knowledge Discovery from Unstructured Data in Financial Services February 7th, 2020 Masaya Abe1 and Junpei Komiyama2 A Robust Transferable Deep Learning Framework for Cross-sectional Investment Strategy 1. Nomura Asset Management Co., Ltd. 2. New York University Kei Nakagawa1, https://arxiv.org/pdf/1910.01491.pdfArXiv version (Full paper):
- 2. 1. Introduction and Motivation 2. Data and Methodology 3. Experimental Results 4. Conclusion Agenda 1
- 3. Agenda 1. Introduction and Motivation 2. Data and Methodology 3. Experimental Results 4. Conclusion 2
- 4. ROE 1 Month Return ● ● ● Score Value Growth Quality Momentum ・Linear Regression Factor Candidates Factor Classification by human Calculate Relative Goodness Cross-sectional Investment Strategy Relative Stock Returns ・Rank IC (Spearman correlation): 3 ・ How average returns change with different stock attributes : Factor
- 5. RIC-NN: Our methodology Score Input Output (3) Deep Transfer Learning Loss rank IC Stop Epoch Time step t-1 Time step t 𝒗𝒊 𝒗 𝒇 (2) Weight Initialization and Stopping Stop 𝒗 𝒇 Loss (1) Multi-factor Deep Learning Approach Score Stock Input Output Score Stock Input Output Source Domain Target Domain Transfer Stock Factor Stock Factor Stock Stock Factor Initialization 𝒗𝒊 NorthAmericaAsiaPacific Initialization 4
- 6. Score Factor Candidates Deep Learning ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ROE 1 Month Return ● ● ● RIC-NN: Multi-factor Deep Learning Approach Calculate Relative Goodness 5 ・ Deep learning for cross sectional investment strategy
- 7. ・ DL for stock return prediction easily overfits to training data. ✓ Use early stopping to control the fitness to the past data. RIC-NN: Weight Initialization and Stopping ・Use RankIC (Spearman correlation) in terms of the fitness. -> intuitive and controllable. Loss rank IC Epoch Time step T-1 Time step T Stop Loss Initialization ・ Epoch-based stopping, Risk of overfitting or underfitting because the training speed varies.0.20.16 Stopping: the rank IC reaches 0.20. Initialization: Use the model of timestep t-1 when rank IC is 0.16. c.f.: fitness of a good portfolio to future return is around 0.10 Initialization Our Proposed (RIC-NN) ✓ Training at time step t: 6
- 8. ScoreRIC-NN Factor Stock Input Output Score Factor Stock Input Output North America Stock Market Asia Pacific Stock Market (Source Domain) (Target Domain) Transfer RIC-NN ・Augment the model using the knowledge of a larger market. ✓ Use transfer learning RIC-NN: Deep Transfer Learning We want to capture the asymmetric structure between the two markets. 7
- 9. Agenda 1. Introduction and Motivation 2. Data and Methodology 3. Experimental Results 4. Conclusion 8
- 10. ・ We use the 20 factors that are often used in practice. ✓ Calculated for each regional index constituents - MSCI North America Index (NA) - MSCI Pacific Index (PA) Features (Various Factors) No. Feature (Factor) No. Feature (Factor) No. Feature (Factor) 1 Book-to-market Ratio 8 Return on Invested Capital 15 EPS Revision(1 month) 2 Earnings-to-price Ratio 9 Accruals 16 EPS Revision(3 months) 3 Dividend Yield 10 Total Asset Growth Rate 17 Past Stock Return(1 month) 4 Sales-to-price Ratio 11 Current Ratio 18 Past Stock Return(12 months) 5 Cash flow-to-price Ratio 12 Equity Ratio 19 Volatility 6 Return on Equity 13 Total Asset Turnover Rate 20 Skewness 7 Return on Asset 14 CAPEX Growth Rate ※ Monthly data Data sources: Namely, Compustat, WorldScope, Thomson Reuters, I/B/E/S and EXSHARE. 9
- 11. Features (Various Factors) ・ Cumulative returns in NA(left-side) and PA(right-side) on 20 factors ・ Return of each factor varies largely over time. 10
- 12. Problem Formulation MSE 𝑡 = 1 𝐾 𝑡′=𝑡−𝑁 𝑡−1 𝑖∈𝑈 𝑡′ 𝑟𝑖,𝑡′+1 − 𝑓 𝒗𝑖,𝑇; 𝜽 𝑇+1 2 𝑁 = 120 (10 years) 𝐾 = 𝑡′=𝑡−𝑁 𝑡−1 𝑈 𝑡′ ・ We define the problem as a regression problem to minimize MSE. ✓ Approximate function 𝑓 ∙ with the parameter 𝜽 𝑇+1 that maps 𝒗𝑖,𝑇 to 𝑟𝑖,𝑇+1 𝑓 𝒗𝑖,𝑇; 𝜽 𝑇+1 → 𝑟𝑖,𝑇+1 ✓ Train the models using the data of the latest 120 time steps from the past 10 years. 11 𝒗𝑖,𝑇 ：Augmented factors 𝜽 𝑇+1：NN Weights 𝑟𝑖,𝑇+1：Relative Stock return
- 13. ✓ 20 factors: Problem Formulation ・ Given a stock 𝑖 at month 𝑇 (𝑖 ∈ 𝑈 𝑇: a regional index constituents at 𝑇) 𝒙𝑖,𝑇 ∈ 𝑅20 ✓ Features: 20 factors and preprocessed factors 𝑥/ 𝑅 𝑦 ≔ 2(𝑥 − 𝑦)/( 𝑥 + 𝑦 ) ✓ Output variable: scaled one-month-ahead stock return Scale to the range [0,1] Pre-processing ＆ Feature augmentation 𝒗𝑖,𝑇 = (𝒙𝑖,𝑇, 𝒙𝑖,𝑇−3, … , 𝒙𝑖,𝑇−12, 𝒙𝑖,𝑇/ 𝑅 𝒙𝑖,𝑇−3, … , 𝒙𝑖,𝑇/ 𝑅 𝒙𝑖,𝑇−12) ∈ [0,1] 𝟏𝟖𝟎 𝑟𝑖,𝑇+1 ∈ [0,1] 12 ・Most of factors are updated quarterly ・Time difference between the present and each quarter ago
- 14. ・ Architecture of RIC-NN is quite standard ✓ Fully-connected feedforward neural networks ✓ 6 Hidden layers: { 150 – 150 – 100 – 100 – 50 – 50 } Dropout rates: (50% – 50% – 30% – 30% – 10% – 10%) ✓ Activation function: ReLU function ✓ RIC-NN(Transfer Learning:TF) Compared Models ・ Other off-the-shelf machine learning models ✓ Epoch-based Neural Network (NN(Epoch)) ✓ Random Forest (RF) ✓ Ridge Regression (RR) We use the weights of the first four layers that are trained in the source region as the initial weight of the target region. Our Proposed (RIC-NN) 13
- 15. Prediction Period November 2004 : December 1994 Scores January 2005 Training 120 set 𝑓 𝒗𝑖,𝑡; 𝜽 𝒕+𝟏 ∗ argmin 𝜽 MSE 𝑡 Features: 𝒗𝑖,𝑡 December 2004 𝒗𝑖,𝑡 𝜽 𝑡+1 ∗ December 2004 : January 1995 February 2005 January 2005 𝒗𝑖,𝑡 𝜽 𝑡+1 ∗ ・・・ October 2018 : November 2008 December 2018 November 2018 𝜽 𝑡+1 ∗𝒗𝑖,𝑡 14 years (168 months) ・ 14 years (from January 2005 to December 2018) ・ Updated by sliding one-month-ahead and carrying out a monthly forecast. 14
- 16. Agenda 1. Introduction and Motivation 2. Data and Methodology 3. Experimental Results 4. Conclusion 15
- 17. ・ Simple portfolio strategies Performance Measure ✓ Long Portfolio Strategy ✓ We make quintile portfolios. - Buy (Long) the top 1/5 score stocks with equal weighting - Benchmark: the average return of all stocks → Relative performance evaluation １ ２ ３ ４ ５ Relative goodness Investment Universe 16
- 18. Performance Measure ・ We use the following (standard) measures. ✓ Alpha Return ≔ ς 𝑡=1 𝑇 1 + 𝛼 𝑡 12/𝑇 − 1 ✓ 𝑇𝐸 Risk ≔ 12 𝑇−1 𝛼 𝑡 − 𝜇 𝛼 2 ✓ 𝐼𝑅 Return/Risk ≔ Alpha/𝑇𝐸 portfolio return – benchmark return 𝜇 𝛼: Average of Alpha 𝛼 𝑡: ✓ MaxDD Worst − case Loss ≔ min 𝑘∈[1,𝑇] (0, 𝑊𝑘 𝑃𝑜𝑟𝑡 max 𝑗∈ 1,𝑘 𝑊𝑗 𝑃𝑜𝑟𝑡 − 1) 𝑊𝑘 𝑃𝑜𝑟𝑡 : Cumulative return of the portfolio 17
- 19. Experimental Results (1/2) ・ NA: RIC-NN without transfer learning performed best. ・ PA: RIC-NN with transfer learning performed best. ・ NA as a source domain enhances the performance of PA, not vice versa. MSCI North America Linear LR RF DL(Epoch) RIC-NN RIC-NN(TF from PF) Alpha 0.62% 0.79% 0.82% 1.23% 1.20% TE 5.40% 5.14% 4.48% 4.14% 4.43% IR 0.11 0.15 0.18 0.30 0.27 MaxDD -21.84% -24.57% -17.41% -14.37% -20.57% Long Nonlinear MSCI Pacific Linear LR RF DL(Epoch) RIC-NN RIC-NN(TF from NA) Alpha 5.35% 3.79% 4.34% 5.25% 5.78% TE 5.17% 5.75% 4.18% 4.20% 3.95% IR 1.04 0.66 1.04 1.25 1.46 MaxDD -11.53% -11.43% -9.37% -7.51% -3.37% Long Nonlinear 18
- 20. Experimental Results (2/2) ・ While NN at epoch 50 performs better in NA, NN at epoch 60 performs better in PA. ・ NN(Epoch) is very sensitive to the choice of the epoch. ・ RIC-NN outperforms epoch-based stopping: rank IC controls the fitness of the stock prediction models consistently. MSCI North America 40 50 56 60 80 Alpha 1.23% 0.18% 1.48% 0.82% 1.25% 0.70% TE 4.14% 4.52% 4.35% 4.48% 4.49% 4.14% IR 0.30 0.04 0.34 0.18 0.28 0.17 MaxDD -14.37% -22.67% -13.48% -17.41% -20.98% -15.94% Long RIC-NN NN(Epoch) MSCI Pacific 40 46 50 60 80 Alpha 5.25% 4.13% 4.34% 4.28% 4.52% 2.99% TE 4.20% 4.36% 4.18% 4.73% 4.34% 4.06% IR 1.25 0.95 1.04 0.90 1.04 0.74 MaxDD -7.51% -8.08% -9.37% -7.16% -7.45% -7.52% NN(Epoch) Long RIC-NN *These epochs are chosen so that the rank IC reaches 0.20 during the training of the first time step. * * 19
- 21. ・ We have proposed a new stock price prediction framework called RIC-NN by introducing three practical ideas: (1) A nonlinear multi-factor approach is better than a linear approach. (2) Rank IC-based stopping outperforms epoch-based stopping. (3) Multi-region transfer learning works well. ・ Better return of the portfolio, better control of the fitness of the model to the past dataset. Conclusion 20