Deep Learning in Finance Summit schedule

Learning to Trade with Q-RL and DQNs

Price changes in financial products are largely random, representing an efficient market, but are often supplemented by salient features that provide additional structure which can be exploited for trading profits. Experienced traders are skilled at identifying such features and deploying profitable exploits. We present some methods for learning such exploits using Q- function based reinforcement learning and DQNs that are trained on simulation models for markets which progress through levels of realism with data provided by generative models that mimic both the randomness and salient features of the actual markets.

David Samuel is a veteran of the financial markets, working in trading and quantitative research for a number of investment banks and proprietary trading firms. He has a PhD in theoretical physics and published research in a number of top academic journals. David has been a lecturer for the MSc in Mathematical Finance at Oxford University and collaborated on the application of machine learning to trading with researchers at Cambridge University. His recent interest has been on the application of reinforcement learning and deep learning methods to trading. David is co-founder of Prediction Machines which develops algorithms for predicting and trading in commercial transactions markets.

Portfolio Management Using Deep Reinforcement Learning

Prediction using supervised learning algorithms for financial time series modelling is hard and converting predictions into actions requires additional naive layer of logic. In this talk I present a financial-model-free Reinforcement Learning framework to provide a deep machine learning solution to the portfolio management problem. A deep neural network is used for generating signals based on historical price data, these signals are fed to a portfolio memory layer that optimizes transaction costs, the deep reinforcement learning agent trains itself using an exhaustive reward function based on long term and short term market performance. This problem is implemented on the cryptocurrency markets and compared with other well known portfolio management frameworks like trend following, mean reversion and equal weighting scheme. I show that the model outperforms the other frameworks but performance is very much dependent on the overall market trend in the long only setup.

Sonam Srivastava is a quantitative trading professional more than 8 years of professional experience in systematic portfolio management and quantitative trading. She is a IIT Kanpur graduate with a Masters in Financial Engineering from Worldquant University. She has worked as a Portfolio Manager at Qplum, applying machine learning & artificial intelligence to automate investment decision making. She has also worked at HSBC as a Quant Researcher and Edelweiss as an Algorithmic Trader. She is an avid researcher in the field of Quantitative Finance and a Registered Investment Advisor.

From unstructured text to knowledge graphs: how to leverage dark data

The banking world is full of dark data: unstructured information that does not live in orderly databases. As we move toward conversational applications, the ability to dig up dark data will become even more critical. While companies say they want a single view of the customer, what they really need is a multi-dimensional view so that every business stakeholder can look at a client’s information from the perspective that matters to them. Deep learning methods such as named entity recognition and co-reference resolution are used to find structure in unstructured data and populate knowledge databases and graphs.

Present Edouard d'Archimbaud is a data scientist in charge of the Data Science and Artificial Intelligence Lab for BNP Paribas CIB. He joined the group in 2016, having previously held different positions focusing on advanced technological research and operational projects in major banking institutions and hedge funds.

Education Edouard graduated from Ecole Polytechnique in Paris, with a specialisation in applied mathematics and computer science. He holds a M.Sc. in Machine Learning and Computer Vision from ENS Cachan.

Predicting Foreign Exchange Market Turbulence using Thick Neural Networks.

Foreign exchange markets are notoriously unpredictable and often expose international businesses and governments to potentially devastating risks. A case for deploying Thick Neural Networks (TNN) as a method to flag periods of exchange turbulence is presented. Following a brief introduction of the phenomena being modelled (exchange rate volatility) and the method deployed a walkthrough will be given on several (neural network) architecture considerations which enhance overall network predictive performance. A novel way for presenting network topology will be shared along with a method for selecting the best parameters whilst evaluating architectures (Teknikeller, 2014). Thick Neural Networks developed are presented with their out-of- sample turbulence flagging abilities and results are benchmarked against orthodox methods.

An inquisitive engineer turned pioneer economist. Awarded his doctorate (WSU, 2013) in developing novel macro-economic models through deploying Artificial Neural Networks and Machine Learning optimization. A professional committed to his corporate career (Lion) whilst lecturing and unit-coordinating masters and undergraduate units in Finance and Economics (Western Sydney University). Eagerly implementing refined techniques to address the sorts of complexities multinational businesses deal with in the areas of Pricing, Optimisation, Finance, Econometrics and Behavioural Economics. Outside of work he is an avid practitioner of Kendo and ambitious gardener who will go out of his way to find a live jazz band performance.

Aspect extraction for opinion mining with a deep convolutional neural network

In this presentation, we present the first deep learning approach to aspect extraction in opinion mining. Aspect extraction is a subtask of sentiment analysis that consists in identifying opinion targets in opinionated text. A deep convolutional neural network has been used to tag each word in opinionated sentences as either aspect or non-aspect word. We also developed a set of linguistic patterns for the same purpose and combined them with the neural network. The resulting ensemble classifier, coupled with a word-embedding model for sentiment analysis, allowed our approach to obtain significantly better accuracy than state-of-the-art methods.

Sandro Cavallari received his BEng in Telecommunication Engineering in 2012 and his MEng in Computer Science in 2015 both from the University of Trento. After finalizing his thesis at the ADSC of Singapore in collaboration with the University of Illinois at Urbana Champaign, he has been awarded the prestigious SINGA scholarship and started his PhD at Nanyang Technological University in 2015 under the supervision of Dr Cambria. His research areas focus on the application of machine learning and natural language processing technique to perform stock market prediction.

Price Returns Prediction Using Recurrent Neural Networks With Long Short Term Memory (LSTM) Units

Traditionally RNNs have been used for text classification and sentiment analysis; in this paper, we extend the idea to price returns prediction. The primary objective of the models will be to accurately predict 5-minute price returns based on multiple price based features. In our network, we use stacked LSTMs instead of MLP architecture. LSTM units are preferred as its architecture helps in solving the vanishing gradient problem and enables in modeling the long-term dependency which is persistent in financial price series data.

Sujit works at Edelweiss securities as a Quantitative Researcher in the Algorithmic Trading Desk where he uses a host of quantitative techniques to create Alpha strategies. Sujit’s work majorly focuses on analyzing and applying ML algorithms for price prediction and environment modeling, which includes creating feature selection and dimensionality reduction models like RFE, Lasso, t-SNE, and PCA. Over the last couple of years, he has focused his attention on using Deep Learning Algorithms for Price predictions at Mid and High-Frequency intervals. Such algorithms included variants of LSTM RNNs, Deep generative Models, Adversarial, and Evolutionary algorithms. Prior to Edelweiss securities, Sujit was working with Morgan Stanley in the Quantitative and Derivatives Strategies team. Sujit has bachelor’s degree in Electronics and Telecommunication engineering from Mumbai University and an MBA from NMIMS.

Deep Autoencoders in Financial Markets

I will be talking about applications of few deep learning algorithms in finance. First, I will talk about autoencoders for dimensionality reduction in relative value strategy. I will illustrate how it is different from conventional methods such as PCA or factor models. I will also address its limitations and things to keep in mind while building autoencoders for dimensionality reduction on financial data.

I will also talk about Generative Adversarial Networks (GAN) models and how they can be used to generate synthetic stock market returns. I will talk about GANs ability to match the fat tail distributions very accurately. I will address the issue of the autocorrelation of returns and how we can use GANs to capture it.

I am cross asset quantitative analyst at Societe Generale and I specialize in machine learning. I have been building quantitative models to forecast/analyze markets from last 10 years. Prior to Societe Generale, I was working as HFT trader at Edelweiss Securities, Mumbai. I use time series analysis, machine learning (deep learning), stochastic calculus, factor analysis and other mathematical techniques to build my models.

I have bachelor’s degree in Electronics and communication from RVCE, Bangalore and master’s degree in Control and Automation from IIT Delhi. I have several hobbies. I like building robots with computer vision, IOT devices and I have used use machine learning techniques to build models to forecast rainfall, analyze genomics data etc.

Price Returns Prediction Using Recurrent Neural Networks With Long Short Term Memory (LSTM) Units

Traditionally RNNs have been used for text classification and sentiment analysis; in this paper, we extend the idea to price returns prediction. The primary objective of the models will be to accurately predict 5-minute price returns based on multiple price based features. In our network, we use stacked LSTMs instead of MLP architecture. LSTM units are preferred as its architecture helps in solving the vanishing gradient problem and enables in modeling the long-term dependency which is persistent in financial price series data.

Sujit works at Edelweiss securities as a Quantitative Researcher in the Algorithmic Trading Desk where he uses a host of quantitative techniques to create Alpha strategies. Sujit’s work majorly focuses on analyzing and applying ML algorithms for price prediction and environment modeling, which includes creating feature selection and dimensionality reduction models like RFE, Lasso, t-SNE, and PCA. Over the last couple of years, he has focused his attention on using Deep Learning Algorithms for Price predictions at Mid and High-Frequency intervals. Such algorithms included variants of LSTM RNNs, Deep generative Models, Adversarial, and Evolutionary algorithms. Prior to Edelweiss securities, Sujit was working with Morgan Stanley in the Quantitative and Derivatives Strategies team. Sujit has bachelor’s degree in Electronics and Telecommunication engineering from Mumbai University and an MBA from NMIMS.

Portfolio Management Using Deep Reinforcement Learning

Prediction using supervised learning algorithms for financial time series modelling is hard and converting predictions into actions requires additional naive layer of logic. In this talk I present a financial-model-free Reinforcement Learning framework to provide a deep machine learning solution to the portfolio management problem. A deep neural network is used for generating signals based on historical price data, these signals are fed to a portfolio memory layer that optimizes transaction costs, the deep reinforcement learning agent trains itself using an exhaustive reward function based on long term and short term market performance. This problem is implemented on the cryptocurrency markets and compared with other well known portfolio management frameworks like trend following, mean reversion and equal weighting scheme. I show that the model outperforms the other frameworks but performance is very much dependent on the overall market trend in the long only setup.

Sonam Srivastava is a quantitative trading professional more than 8 years of professional experience in systematic portfolio management and quantitative trading. She is a IIT Kanpur graduate with a Masters in Financial Engineering from Worldquant University. She has worked as a Portfolio Manager at Qplum, applying machine learning & artificial intelligence to automate investment decision making. She has also worked at HSBC as a Quant Researcher and Edelweiss as an Algorithmic Trader. She is an avid researcher in the field of Quantitative Finance and a Registered Investment Advisor.

Despite the fact that the topic of predictive scoring models building is often considered in articles and it was written not one book about it, many things remain out of sight. An attempt to consider some of interesting questions will be made in the presentation. It will be practical ways which help to answer on such questions as: how to estimate the maximum Gini before the start of the model building; how we could add non-maturing loans in likelihood error function; one additional way how we could estimate overfitting and how we can understand need we separate Fraud Prevention Model or not.

Alexey Pogrebnyak is a head of Decision Science department in financial startup Finstar which was launched in APAC two years ago. Alexey graduated from Moscow State University, with a specialization in high energy physics.

The usage of machine learning techniques for the prediction of financial time series has been used investigated for a (very) long time. Both in the academic and fund management words, generative methods such as Switching Autoregressive Hidden Markov and changepoint models are generally found to be unsuccessful at predicting daily (and higher frequency) prices from a wide range of asset classes. However, committees of discriminative techniques (such as Support Vector Machines or Relevance Vector Machines and Neural Networks) are found to give some interesting results. It however remains that there is no such thing as a magic recipe and that using machine learning model for financial markets faces enormous challenge due to the very nature of financial markets, i.e. an ecology of learning agents continuously creating information and having different

Joel Guglietta has been working as a quantitative strategist & portfolio manager for hedge funds and investment banks (Brevan Howard, BTIM, Graticule, HSBC) for more than 12 years. His expertise is in quantitative models for asset allocation, portfolio construction & management using a wide range of technics of which machine learning techniques and genetic algorithms.

Transforming Insurance through Machine Learning

Machine Learning and AI is playing crucial in transforming traditional insurers. It is helping insurers to deliver high value to its customers by simplifying products, standardizing underwriting, designing on-demand pricing solutions (telematics), and identifying fraudulent claims etc. In the presentation, I would share application of machine learning in insurance, insights about dealing with finance data , feature engineering and highlight the usability of machine learning models on real business problems.

Siddhant is an analytics professional working as a Data Scientist at AXA Data Innovation Lab, Singapore. His current role involves executing data science projects to support AXA entities in sales and marketing activities such as opportunity sizing, cross-sell and up-sell, product recommendations, quantifying risk, retention and fraud using advanced machine learning techniques. Previously, at ZS Associates, he consulted many Fortune 500 pharmaceutical clients in devising sales and marketing strategies to improve brand positioning in the market using data analytics. He holds a Bachelors of Engineering (Hons.) in Electrical and Electronics from BITS Pilani, India and MSc in Business Analytics from National University of Singapore.

Deep learning for stock market prediction

In this talk I share with you two of our recent research projects leveraging unstructured data for predicting stock market movements. In the first project, we use open information extraction techniques to extract event structure out of news titles, representing the results using deep learning, and building neural network models to match them with S&P 500 company returns. In the second project, we extend the efforts further, by using deep learning models to directly learn representations of news abstracts, from which we predict cumulative abnormal returns of S&P companies over 3 days. The results are promising, beating sentiment baselines significantly without using time series data.

Yue Zhang is currently an assistant professor at Singapore University of Technology and Design. Before joining SUTD in July 2012, he worked as a postdoctoral research associate in University of Cambridge, UK. Yue Zhang received his DPhil and MSc degrees from University of Oxford, UK, and his BEng degree from Tsinghua University, China. His research interests include natural language processing, machine learning and artificial Intelligence. He has been working on statistical parsing, parsing, text synthesis, machine translation, sentiment analysis and stock market analysis intensively. Yue Zhang serves as the reviewer for top journals such as Computational Linguistics, Transaction of Association of Computational Linguistics and Journal of Artificial Intelligence Research. He is also PC member for conferences such as ACL, COLING, EMNLP, NAACL, EACL, AAAI and IJCAI. Recently, he was the area chairs of COLING 2014, NAACL 2015, EMNLP 2015, ACL 2017 and EMNLP2017.

Deep Reinforcement Learning for Optimal Order Placement in a Limit Order Book

Financial trading is essentially a search problem. The buy-side agent must find a counterpart sell-side agent willing to trade the financial asset at the set quantity and price. The virtual space where the agents execute their trading actions is called limit-order book. We present a deep reinforcement learning algorithm for optimizing the execution of limit-order actions to find an optimal order placement. The reinforcement learning agent utilizes historical limit-order data to learn to an optimal compromise between fast order completion but with higher costs and slow, riskier order completion but with lower costs. We also give a technological overview of the system and discuss the challenges and potential future work.

Ph.D. candidate in Deep Learning, M.Eng. in Software Engineering for Machine Learning.

He is Interested in: Multimodal Deep Learning; Non-convex Optimization; (Visual) Question Answering; Natural Language Processing and Generation.

Investment Selection By Combining Chaos Theory with Artificial Intelligence

The I Know First forecasting algorithm models the markets as non-stationary chaotic systems with fractal properties. Features: Multi-representation – wide and deep supervised learning, data munging to expose important features. Generalization through explicit regularization. Plurality of forecasts. Principal Component Analysis results in near orthogonal inputs. Genetic adaptable algorithm. Can Reinforcement Learning system learn to trade the market? We will report here our results of using Reinforcement Learning to create a trading system based on the I Know First forecasts.
With over 35 years of research in artificial intelligence and machine learning, Dr. Lipa Roitman has developed algorithms to predict chaotic events in financial markets. He has a M.Sc. in chemistry and organic chemistry from Novosibirsk University, a Ph.D. in organic and physical organic chemistry, and photochemistry from the Weizmann Institute of Science, a post doctorate in polymer chemistry from the University of Akron, and has a background in mathematics and statistics. His past work includes: R&D Chemist for Rohm and Haas Corp. USA.

As Co-Founder and CTO of I Know First Ltd., the company uses his algorithm to predict markets and supply big data solutions for institutional and large private investors. Lipa is responsible for the design, development, characterization, specification, and implementation of algorithmic trading solutions. In addition, his other responsibilities include development of quantitative trading strategies, predictive analytics, and development of quantitative trading strategies, algorithmic trading applications, and Big-Data solutions for hedge funds and institutional clients.

Investment Selection By Combining Chaos Theory with Artificial Intelligence

Yaron is the CEO of I Know First, an Israeli AI-Wealth Tech company he co-founded in 2011. I Know First specializes on the development and application of advanced self-learning forecasting algorithms for the capital markets, utilizing artificial intelligence and machine learning. Predictions are generated for over 10,000 securities and are applied to discover best investment opportunities, support existing research and investment processes as well as structure AI-powered investment products. The concept of the current algorithm has crystallized following decades of prior research into the nature of chaotic systems by I Know First's R&D team. The algorithm incorporates multi-layered neural networks and genetic algorithms, it is adaptable and scalable allowing comprehensive customized algorithmic solutions depending on clients' needs - banks, asset management first, family offices, hedge funds and self-directed DIY-investors.

After completing his degree in Engineering at the Tel-Aviv University, Yaron worked as a division manager at OIC, one of the Israeli leading consulting firms. Yaron Golgher has over 18 years of experience leading AI, deep learning and machine learning projects including development of AI forecasting algorithms and predictive analytics, quantitative trading strategies, algorithmic trading applications, and big-data solutions for hedge funds and institutional clients.

He received his executive MBA from Ben Gurion University, before founding I Know First.

Rasmi is a former computational biology researcher, software programmer, data scientist transformed to a Technologist. Currently one of key members of OCBC Bank’s FinTech & Innovation Group: The Open Vault, where he tries to filter noise from hype. He enjoys connecting diverse dots of any use case and bring an idea to implementation. In his free time, he likes to play around with Raspberry Pi based Home Automation System. His current areas of interest include: blockchain, chatbots, core banking 2.0 and last mile supply chain management using IoT.

Varun Mittal is the ASEAN Fintech Lead at EY covering emerging technology and business models in financial services across ASEAN. Previously Varun was the first marketing and sales employee at helloPay (acquired by ANT Financial, Alibaba Group), led payments for Samsung in ASEAN and developed regional mobile payment solutions at Singtel Group. Prior his work in payments, Mittal worked at Microsoft and Gemalto. He holds an MBA from the National University of Singapore. Varun is also part of founding team of Singapore Fintech Association, working closely with startups, educational institutions, investors and regulators across ASEAN.

Gabriel Lundberg is an operating partner at SeedPlus, a venture capital fund focused on operational involvement where applied machine learning is a fundamental pillar of the investment philosophy. A reformed engineering physicist with graduate studies in adaptive systems, Gabriel previously spent six years at Spotify as a product manager. When not dealing with product strategy or user experience design, he still claims to be able to write code in a pinch.