Martino Bernasconi, S. Martino, Edoardo Vittori, F. Trovò, Marcello Restelli
We study the problem of developing a Smart Order Routing algorithm that learns how to optimize the dollar volume, i.e., the total value of the traded shares, gained from slicing an order across multiple dark pools. Our work is motivated by two distinct issues: (i) the surge in liquidity fragmentation caused by the rising popularity of electronic trading and by the increasing number of trading venues, and (ii) the growth in popularity of dark pools, an exchange venue characterised by a lack of transparency. This paper critically discusses the known dark pool literature and proposes a novel algorithm, namely the DP-CMAB algorithm, that extends existing solutions by allowing the agent to specify the desired limit price when placing orders. Specifically, we frame the problem of dollar volume optimization in a multi-venue setting as a Combinatorial Multi-Armed Bandit (CMAB) problem, representing a generalization of the well-studied MAB framework. Drawing from the rich MAB and CMAB literature, we present multiple strategies that our algorithm may adopt to select the best allocation options. Furthermore, we analyze how exploiting financial domain knowledge improves the agents’ performance. Finally, we evaluate the DP-CMAB performance in an environment built from real market data and show that our algorithm outperforms state-of-the-art solutions.
{"title":"Dark-Pool Smart Order Routing: a Combinatorial Multi-armed Bandit Approach","authors":"Martino Bernasconi, S. Martino, Edoardo Vittori, F. Trovò, Marcello Restelli","doi":"10.1145/3533271.3561728","DOIUrl":"https://doi.org/10.1145/3533271.3561728","url":null,"abstract":"We study the problem of developing a Smart Order Routing algorithm that learns how to optimize the dollar volume, i.e., the total value of the traded shares, gained from slicing an order across multiple dark pools. Our work is motivated by two distinct issues: (i) the surge in liquidity fragmentation caused by the rising popularity of electronic trading and by the increasing number of trading venues, and (ii) the growth in popularity of dark pools, an exchange venue characterised by a lack of transparency. This paper critically discusses the known dark pool literature and proposes a novel algorithm, namely the DP-CMAB algorithm, that extends existing solutions by allowing the agent to specify the desired limit price when placing orders. Specifically, we frame the problem of dollar volume optimization in a multi-venue setting as a Combinatorial Multi-Armed Bandit (CMAB) problem, representing a generalization of the well-studied MAB framework. Drawing from the rich MAB and CMAB literature, we present multiple strategies that our algorithm may adopt to select the best allocation options. Furthermore, we analyze how exploiting financial domain knowledge improves the agents’ performance. Finally, we evaluate the DP-CMAB performance in an environment built from real market data and show that our algorithm outperforms state-of-the-art solutions.","PeriodicalId":134888,"journal":{"name":"Proceedings of the Third ACM International Conference on AI in Finance","volume":"114 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-10-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"124771011","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Market making strategy is one of the most popular high frequency trading strategies, where a market maker continuously quotes on both bid and ask side of the limit order book to profit from capturing bid-ask spread and to provide liquidity to the market. A market maker should consider three types of risk: 1) inventory risk, 2) adverse selection risk, and 3) non-execution risk. While there have been a lot of studies on market making via deep reinforcement learning, most of them focus on the first risk. However, in highly competitive markets, the latter two risks are very important to make stable profit from market making. For better control of the latter two risks, it is important to reserve good queue position of their resting limit orders. For this purpose, practitioners frequently adopt order stacking framework where their limit orders are quoted at multiple price levels beyond the best limit price. To the best of our knowledge, there have been no studies that adopt order stacking framework for market making. In this regard, we develop a deep reinforcement learning model for market making under order stacking framework. We use a modified state representation to efficiently encode the queue positions of the resting limit orders. We conduct comprehensive ablation study to show that by utilizing deep reinforcement learning, a market making agent under order stacking framework successfully learns to improve the PL while reducing various risks. For the training and testing of our model, we use complete limit order book data of KOSPI200 Index Futures from November 1, 2019 to January 31, 2020 which is comprised of 61 trading days.
{"title":"Market Making under Order Stacking Framework: A Deep Reinforcement Learning Approach","authors":"G. Chung, Munki Chung, Yongjae Lee, W. Kim","doi":"10.1145/3533271.3561789","DOIUrl":"https://doi.org/10.1145/3533271.3561789","url":null,"abstract":"Market making strategy is one of the most popular high frequency trading strategies, where a market maker continuously quotes on both bid and ask side of the limit order book to profit from capturing bid-ask spread and to provide liquidity to the market. A market maker should consider three types of risk: 1) inventory risk, 2) adverse selection risk, and 3) non-execution risk. While there have been a lot of studies on market making via deep reinforcement learning, most of them focus on the first risk. However, in highly competitive markets, the latter two risks are very important to make stable profit from market making. For better control of the latter two risks, it is important to reserve good queue position of their resting limit orders. For this purpose, practitioners frequently adopt order stacking framework where their limit orders are quoted at multiple price levels beyond the best limit price. To the best of our knowledge, there have been no studies that adopt order stacking framework for market making. In this regard, we develop a deep reinforcement learning model for market making under order stacking framework. We use a modified state representation to efficiently encode the queue positions of the resting limit orders. We conduct comprehensive ablation study to show that by utilizing deep reinforcement learning, a market making agent under order stacking framework successfully learns to improve the PL while reducing various risks. For the training and testing of our model, we use complete limit order book data of KOSPI200 Index Futures from November 1, 2019 to January 31, 2020 which is comprised of 61 trading days.","PeriodicalId":134888,"journal":{"name":"Proceedings of the Third ACM International Conference on AI in Finance","volume":"10 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-10-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129888118","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Semi supervised learning has attracted attention of AI researchers in the recent past, especially after the advent of deep learning methods and their success in several real world applications. Most deep learning models require large amounts of labelled data, which is expensive to obtain. Fraud detection is a very important problem for several industries and large amount of data is often available. However, obtaining labelled data is cumbersome and hence semi-supervised learning is perfectly positioned to aid us in building robust and accurate supervised models. In this work, we consider different kinds of fraud detection paradigms and show that a self-training based semi-supervised learning approach can produce significant improvements over a model that has been training on a limited set of labelled data. We propose a novel self-training approach by using a guided sharpening technique using a pair of autoencoders which provide useful cues for incorporating unlabelled data in the training process. We conduct thorough experiments on three different real world databases and analysis to showcase the effectiveness of the approach. On the elliptic bitcoin fraud dataset, we show that utilizing unlabelled data improves the F1 score of the model trained on limited labelled data by around 10%.
{"title":"Guided Self-Training based Semi-Supervised Learning for Fraud Detection","authors":"Awanish Kumar, Soumyadeep Ghosh, Janu Verma","doi":"10.1145/3533271.3561783","DOIUrl":"https://doi.org/10.1145/3533271.3561783","url":null,"abstract":"Semi supervised learning has attracted attention of AI researchers in the recent past, especially after the advent of deep learning methods and their success in several real world applications. Most deep learning models require large amounts of labelled data, which is expensive to obtain. Fraud detection is a very important problem for several industries and large amount of data is often available. However, obtaining labelled data is cumbersome and hence semi-supervised learning is perfectly positioned to aid us in building robust and accurate supervised models. In this work, we consider different kinds of fraud detection paradigms and show that a self-training based semi-supervised learning approach can produce significant improvements over a model that has been training on a limited set of labelled data. We propose a novel self-training approach by using a guided sharpening technique using a pair of autoencoders which provide useful cues for incorporating unlabelled data in the training process. We conduct thorough experiments on three different real world databases and analysis to showcase the effectiveness of the approach. On the elliptic bitcoin fraud dataset, we show that utilizing unlabelled data improves the F1 score of the model trained on limited labelled data by around 10%.","PeriodicalId":134888,"journal":{"name":"Proceedings of the Third ACM International Conference on AI in Finance","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-10-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"130357306","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Counterfactual explanation has been the core of interpretable machine learning, which requires a trained model to be able to not only infer but also justify its inference. This problem is crucial in many fields, such as fintech and the healthcare industry, where accurate decisions and their justifications are equally important. Many studies have leveraged the power of deep generative models for counterfactual generation. However, most focus on vision data and leave the latent space unsupervised. In this paper, we propose a new and general framework that uses a supervised extension to the Variational Auto-Encoder (VAE) with Normalizing Flow (NF) for simultaneous classification and counterfactual generation. We show experiments on two tabular financial data-sets, Lending Club (LCD) and Give Me Some Credit (GMC), which show that the model can achieve a state-of-art level prediction accuracy while also producing meaningful counterfactual examples to interpret and justify the classifier’s decision.
反事实解释一直是可解释性机器学习的核心，这需要一个训练有素的模型不仅能够推断，而且能够证明其推断。这个问题在许多领域至关重要，比如金融科技和医疗保健行业，在这些领域，准确的决策及其理由同样重要。许多研究利用深度生成模型的力量来生成反事实。然而，大多数研究都集中在视觉数据上，留下了不受监督的潜在空间。在本文中，我们提出了一个新的和通用的框架，该框架使用了具有归一化流(NF)的变分自编码器(VAE)的监督扩展，用于同时分类和反事实生成。我们展示了两个表格金融数据集的实验，Lending Club (LCD)和Give Me Some Credit (GMC)，这表明该模型可以达到最先进的预测精度，同时也产生了有意义的反事实示例来解释和证明分类器的决定。
{"title":"An Interpretable Deep Classifier for Counterfactual Generation","authors":"Wei Zhang, Brian Barr, J. Paisley","doi":"10.1145/3533271.3561722","DOIUrl":"https://doi.org/10.1145/3533271.3561722","url":null,"abstract":"Counterfactual explanation has been the core of interpretable machine learning, which requires a trained model to be able to not only infer but also justify its inference. This problem is crucial in many fields, such as fintech and the healthcare industry, where accurate decisions and their justifications are equally important. Many studies have leveraged the power of deep generative models for counterfactual generation. However, most focus on vision data and leave the latent space unsupervised. In this paper, we propose a new and general framework that uses a supervised extension to the Variational Auto-Encoder (VAE) with Normalizing Flow (NF) for simultaneous classification and counterfactual generation. We show experiments on two tabular financial data-sets, Lending Club (LCD) and Give Me Some Credit (GMC), which show that the model can achieve a state-of-art level prediction accuracy while also producing meaningful counterfactual examples to interpret and justify the classifier’s decision.","PeriodicalId":134888,"journal":{"name":"Proceedings of the Third ACM International Conference on AI in Finance","volume":"124 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-10-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116325405","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
We develop a simple and computationally efficient significance test for the features of a predictive model. Our forward-selection approach applies to any model specification, learning task and variable type. The test is non-asymptotic, straightforward to implement, and does not require model refitting. It identifies the statistically significant features as well as feature interactions of any order in a hierarchical manner, and generates a model-free notion of feature importance. This testing procedure can be used for model and variable selection. Experimental and empirical results illustrate its performance.
{"title":"Computationally Efficient Feature Significance and Importance for Predictive Models","authors":"Enguerrand Horel, K. Giesecke","doi":"10.1145/3533271.3561713","DOIUrl":"https://doi.org/10.1145/3533271.3561713","url":null,"abstract":"We develop a simple and computationally efficient significance test for the features of a predictive model. Our forward-selection approach applies to any model specification, learning task and variable type. The test is non-asymptotic, straightforward to implement, and does not require model refitting. It identifies the statistically significant features as well as feature interactions of any order in a hierarchical manner, and generates a model-free notion of feature importance. This testing procedure can be used for model and variable selection. Experimental and empirical results illustrate its performance.","PeriodicalId":134888,"journal":{"name":"Proceedings of the Third ACM International Conference on AI in Finance","volume":"91 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-10-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123173400","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
D. Zhou, Ajim Uddin, Zuofeng Shang, C. Sylla, Dantong Yu
Many finance time-series analyses often track a matrix of variables at each time and study their co-evolution over a long time. The matrix time series is overly sparse, involves complex interactions among latent matrix factors, and demands advanced models to extract dynamic temporal patterns from these interactions. This paper proposes a Core Matrix Regression with Regularization algorithm (CMRR) to capture spatiotemporal relations in sparse matrix-variate time series. The model decomposes each matrix into three factor matrices of row entities, column entities, and interactions between row entities and column entities, respectively. Subsequently, it applies recurrent neural networks on interaction matrices to extract temporal patterns. Given the sparse matrix, we design an element-wise orthogonal matrix factorization that leverages the Stochastic Gradient Descent (SGD) in a deep learning platform to overcome the challenge of the sparsity and large volume of complex data. The experiment confirms that combining orthogonal matrix factorization with recurrent neural networks is highly effective and outperforms existing graph neural networks and tensor-based time series prediction methods. We apply CMRR in three real-world financial applications: firm earning forecast, predicting firm fundamentals, and firm characteristics, and demonstrate its consistent performance superiority: reducing error by 23%-53% over other state-of-the-art high-dimensional time series prediction algorithms.
{"title":"Core Matrix Regression and Prediction with Regularization","authors":"D. Zhou, Ajim Uddin, Zuofeng Shang, C. Sylla, Dantong Yu","doi":"10.1145/3533271.3561709","DOIUrl":"https://doi.org/10.1145/3533271.3561709","url":null,"abstract":"Many finance time-series analyses often track a matrix of variables at each time and study their co-evolution over a long time. The matrix time series is overly sparse, involves complex interactions among latent matrix factors, and demands advanced models to extract dynamic temporal patterns from these interactions. This paper proposes a Core Matrix Regression with Regularization algorithm (CMRR) to capture spatiotemporal relations in sparse matrix-variate time series. The model decomposes each matrix into three factor matrices of row entities, column entities, and interactions between row entities and column entities, respectively. Subsequently, it applies recurrent neural networks on interaction matrices to extract temporal patterns. Given the sparse matrix, we design an element-wise orthogonal matrix factorization that leverages the Stochastic Gradient Descent (SGD) in a deep learning platform to overcome the challenge of the sparsity and large volume of complex data. The experiment confirms that combining orthogonal matrix factorization with recurrent neural networks is highly effective and outperforms existing graph neural networks and tensor-based time series prediction methods. We apply CMRR in three real-world financial applications: firm earning forecast, predicting firm fundamentals, and firm characteristics, and demonstrate its consistent performance superiority: reducing error by 23%-53% over other state-of-the-art high-dimensional time series prediction algorithms.","PeriodicalId":134888,"journal":{"name":"Proceedings of the Third ACM International Conference on AI in Finance","volume":"4 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-10-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"114208993","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
An investor short sells when he/she borrows a security and sells it on the open market, planning to buy it back later at a lower price. That said, short-sellers profit from a drop in the price of the security. If the shares of the security instead increase in price, short sellers can bare large losses. Short interest stock market data, provide crucial information of short selling in the market for data mining by publishing the number of shares that have been sold short. Short interest reports are compiled and published by the regulators at a high cost. In particular, brokers and market participants must report their positions on a daily basis to Financial Industry Regulatory Authority (FINRA). Then, FINRA processes the data and provides aggregated feeds to potential clients at a high cost. Third party data providers offer the same service at a lower cost given that the brokers contribute their data to the aggregated data feeds. However, the aggregated feeds do not cover 100% of the market since the brokers are not willing to submit and trust their individual data with the data providers. Not to mention that brokers and market participants do not wish to reveal such information on a daily basis to a third party. In this paper, we show how to publish short interest stock market data using Secure Multiparty Computation: In our process, brokers and market participants submit to a data provider their short selling information, including the symbol of the security and its volume in encrypted messages on a daily basis. The messages are encrypted in a way that the data provider cannot decrypt them and therefore cannot learn about individual participants input. Then, the data provider, can compute an aggregation on the encrypted data and publish the aggregation of the volume per security. It is important to note that the individual volumes are not revealed to the data provider, only the aggregated volume is published.
{"title":"Addressing Extreme Market Responses Using Secure Aggregation","authors":"Sahar Mazloom, Antigoni Polychroniadou, T. Balch","doi":"10.1145/3533271.3561776","DOIUrl":"https://doi.org/10.1145/3533271.3561776","url":null,"abstract":"An investor short sells when he/she borrows a security and sells it on the open market, planning to buy it back later at a lower price. That said, short-sellers profit from a drop in the price of the security. If the shares of the security instead increase in price, short sellers can bare large losses. Short interest stock market data, provide crucial information of short selling in the market for data mining by publishing the number of shares that have been sold short. Short interest reports are compiled and published by the regulators at a high cost. In particular, brokers and market participants must report their positions on a daily basis to Financial Industry Regulatory Authority (FINRA). Then, FINRA processes the data and provides aggregated feeds to potential clients at a high cost. Third party data providers offer the same service at a lower cost given that the brokers contribute their data to the aggregated data feeds. However, the aggregated feeds do not cover 100% of the market since the brokers are not willing to submit and trust their individual data with the data providers. Not to mention that brokers and market participants do not wish to reveal such information on a daily basis to a third party. In this paper, we show how to publish short interest stock market data using Secure Multiparty Computation: In our process, brokers and market participants submit to a data provider their short selling information, including the symbol of the security and its volume in encrypted messages on a daily basis. The messages are encrypted in a way that the data provider cannot decrypt them and therefore cannot learn about individual participants input. Then, the data provider, can compute an aggregation on the encrypted data and publish the aggregation of the volume per security. It is important to note that the individual volumes are not revealed to the data provider, only the aggregated volume is published.","PeriodicalId":134888,"journal":{"name":"Proceedings of the Third ACM International Conference on AI in Finance","volume":"207 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-10-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115562100","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Artificial intelligence-based algorithmic trading has recently started to attract more attention. Among the techniques, deep learning-based methods such as transformers, convolutional neural networks, and patch embedding approaches have become quite popular inside the computer vision researchers. In this research, inspired by the state-of-the-art computer vision methods, we have come up with 2 approaches: DAPP (Deep Attention-based Price Prediction) and DPPP (Deep Patch-based Price Prediction) that are based on vision transformers and patch embedding-based convolutional neural networks respectively to predict asset price and direction from historical price data by capturing the image properties of the historical time-series dataset. Before applying attention-based architecture, we have transformed historical time series price dataset into two-dimensional images by using various number of different technical indicators. Each indicator creates data for a fixed number of days. Thus, we construct two-dimensional images of various dimensions. Then, we use original images valleys and hills to label each image as Hold, Buy, or Sell. We find our trained attention-based models to frequently provide better results for ETFs in comparison to the baseline convolutional architectures in terms of both accuracy and financial analysis metrics during longer testing periods. Our code and processed datasets are available at https://github.com/seferlab/SPDPvCNN
{"title":"Asset Price and Direction Prediction via Deep 2D Transformer and Convolutional Neural Networks","authors":"Tuna Tuncer, Uygar Kaya, Emre Sefer, Onur Alacam, Tugcan Hoser","doi":"10.1145/3533271.3561738","DOIUrl":"https://doi.org/10.1145/3533271.3561738","url":null,"abstract":"Artificial intelligence-based algorithmic trading has recently started to attract more attention. Among the techniques, deep learning-based methods such as transformers, convolutional neural networks, and patch embedding approaches have become quite popular inside the computer vision researchers. In this research, inspired by the state-of-the-art computer vision methods, we have come up with 2 approaches: DAPP (Deep Attention-based Price Prediction) and DPPP (Deep Patch-based Price Prediction) that are based on vision transformers and patch embedding-based convolutional neural networks respectively to predict asset price and direction from historical price data by capturing the image properties of the historical time-series dataset. Before applying attention-based architecture, we have transformed historical time series price dataset into two-dimensional images by using various number of different technical indicators. Each indicator creates data for a fixed number of days. Thus, we construct two-dimensional images of various dimensions. Then, we use original images valleys and hills to label each image as Hold, Buy, or Sell. We find our trained attention-based models to frequently provide better results for ETFs in comparison to the baseline convolutional architectures in terms of both accuracy and financial analysis metrics during longer testing periods. Our code and processed datasets are available at https://github.com/seferlab/SPDPvCNN","PeriodicalId":134888,"journal":{"name":"Proceedings of the Third ACM International Conference on AI in Finance","volume":"5 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-10-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126591083","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
We propose a new portfolio optimization framework, partially egalitarian portfolio selection (PEPS). Inspired by the celebrated LASSO regression, we regularize the mean-variance portfolio optimization by adding two regularizing terms that essentially zero out portfolio weights of some of the assets in the portfolio and select and shrink the portfolio weights of the remaining assets towards the equal weights to hedge against parameter estimation risk. We solve our PEPS formulations by applying recent advances in mixed integer optimization that allow us to tackle large-scale portfolio problems. We also build a predictive regression model for expected return using two cross-sectional factors, the short-term reversal factor and the medium-term momentum factor, that are shown to be the more significant predictive factors among the hundreds of factors tested in the empirical finance literature. We then incorporate our predictive regression into PEPS by replacing the historical mean. We test our PEPS formulations against an array of classical portfolio optimization strategies on a number of datasets in the US equity markets. The PEPS portfolios enhanced with the predictive regression estimates of the expected stock returns exhibit the highest out-of-sample Sharpe ratios in all instances.
{"title":"Portfolio Selection: A Statistical Learning Approach","authors":"Yiming Peng, V. Linetsky","doi":"10.1145/3533271.3561707","DOIUrl":"https://doi.org/10.1145/3533271.3561707","url":null,"abstract":"We propose a new portfolio optimization framework, partially egalitarian portfolio selection (PEPS). Inspired by the celebrated LASSO regression, we regularize the mean-variance portfolio optimization by adding two regularizing terms that essentially zero out portfolio weights of some of the assets in the portfolio and select and shrink the portfolio weights of the remaining assets towards the equal weights to hedge against parameter estimation risk. We solve our PEPS formulations by applying recent advances in mixed integer optimization that allow us to tackle large-scale portfolio problems. We also build a predictive regression model for expected return using two cross-sectional factors, the short-term reversal factor and the medium-term momentum factor, that are shown to be the more significant predictive factors among the hundreds of factors tested in the empirical finance literature. We then incorporate our predictive regression into PEPS by replacing the historical mean. We test our PEPS formulations against an array of classical portfolio optimization strategies on a number of datasets in the US equity markets. The PEPS portfolios enhanced with the predictive regression estimates of the expected stock returns exhibit the highest out-of-sample Sharpe ratios in all instances.","PeriodicalId":134888,"journal":{"name":"Proceedings of the Third ACM International Conference on AI in Finance","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-10-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"130342375","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
We conduct an extensive empirical analysis to evaluate the performance of the recently developed reinforcement learning algorithms by Jia and Zhou [11] in asset allocation tasks. We propose an efficient implementation of the algorithms in a dynamic mean-variance portfolio selection setting. We compare it with the conventional plug-in estimator and two state-of-the-art deep reinforcement learning algorithms, deep deterministic policy gradient (DDPG) and proximal policy optimization (PPO), with both simulated and real market data. On both data sets, our algorithm significantly outperforms the others. In particular, using the US stocks data from Jan 2000 to Dec 2019, we demonstrate the effectiveness of our algorithm in reaching the target return and maximizing the Sharpe ratio for various periods under consideration, including the period of the financial crisis in 2007-2008. By contrast, the plug-in estimator performs poorly on real data sets, and PPO performs better than DDPG but still has lower Sharpe ratio than the market. Our algorithm also outperforms two well-diversified portfolios: the market and equally weighted portfolios.
{"title":"Achieving Mean–Variance Efficiency by Continuous-Time Reinforcement Learning","authors":"Yilie Huang, Yanwei Jia, X. Zhou","doi":"10.1145/3533271.3561760","DOIUrl":"https://doi.org/10.1145/3533271.3561760","url":null,"abstract":"We conduct an extensive empirical analysis to evaluate the performance of the recently developed reinforcement learning algorithms by Jia and Zhou [11] in asset allocation tasks. We propose an efficient implementation of the algorithms in a dynamic mean-variance portfolio selection setting. We compare it with the conventional plug-in estimator and two state-of-the-art deep reinforcement learning algorithms, deep deterministic policy gradient (DDPG) and proximal policy optimization (PPO), with both simulated and real market data. On both data sets, our algorithm significantly outperforms the others. In particular, using the US stocks data from Jan 2000 to Dec 2019, we demonstrate the effectiveness of our algorithm in reaching the target return and maximizing the Sharpe ratio for various periods under consideration, including the period of the financial crisis in 2007-2008. By contrast, the plug-in estimator performs poorly on real data sets, and PPO performs better than DDPG but still has lower Sharpe ratio than the market. Our algorithm also outperforms two well-diversified portfolios: the market and equally weighted portfolios.","PeriodicalId":134888,"journal":{"name":"Proceedings of the Third ACM International Conference on AI in Finance","volume":"56 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-10-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"130513763","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：