All students are expected to do an internship. The Fin-ML CREATE program facilitates students to find internships through its network of professionals, and access to financial institutions. All internships must be of applied research nature. Students are also encouraged to find their own internships, through their own initiatives or the help of their supervisor, subject to approval by the Fin-ML internship committee. At the end of the internship, the student must present a project report summarizing their work, and be available to present in front of the CREATE community (sensitive data and confidential information will be respected) through either:

  • A report summarizing the research output
  • A conference presentation based on the results
  • An academic paper targeting a submission to a research journal
  • Model code transferred to the paper

Application of GPT to automate financial newsletter production

Date : December 3, 2023

Finliti wants to use artificial intelligence to optimize its business operations. One task they want to optimize is the research, writing and editing of a newsletter that is sent out to the company's mailing list every week. Manually writing this letter on average, according to the writer's experience, takes about 5 hours.

Finliti wanted to investigate to what extent using code to automatically generate the newsletter would reduce the time necessary to produce a newsletter of a similar level of quality. The project to determine this consisted of writing the code, testing it out in practice, and revising it to determine the final improvement over manual writing.

Overall, the amount of time needed dropped by about one hour. One suggested further step to reduce the revision time could be to have the code use the complete text from fewer articles if it can be upgraded with internet access.

By Gavin OROK, currently a Master's candidate in Quantitative Finance at the University of Waterloo. Gavin completed his Undergraduate degree in Mathematical Finance at the same university.
For his Fin-ML internship, he worked at the Toronto-based startup Finliti which intends to educate young professionals on how to invest beyond a passive index fund strategy. He is also a recipient of Fin-ML CREATE grant supervised by prof. Christiane Lemieux.

Classifier Rank - A new model validation method to imbalanced data

Date : November 25, 2022

Most of the commonly used confusion matrix-based classification performance metrics, such as f1_score, MCC, PRC and lift curve, are sensitive to the class imbalance. To address this problem, we propose a novel classifier evaluation method, called classifier rank, which eliminates the effect of class imbalance by providing the rank of the classifier in the space of all possible classifiers.

To rank a classifier, we find the distribution of performance metrics conditional on arbitrary imbalance rate. However, some metrics like PRC and lift curve are functions of a large sequence of confusion matrices whose joint distribution is difficult to estimate.

Hence, we propose a discrete hidden Markov model and a continuous directed binary tree model to effectively represent this large-scale joint distribution. As a result, we can estimate the classifier rank using graphical inference algorithms, such as Variable Elimination and Monte-Carlo algorithm.

By Ningsheng ZHAO, currently a Ph.D. candidate in Information and Systems Engineering from Concordia University. Ningsheng obtained his master's diploma in Statistics from the University of Waterloo. His research interests are ML and Data Science. His current research is focusing on machine Learning models’ diagnostics and explanation techniques and their applications in business, which is joint research with Dr. Jia Yuan Yu and Dr. Krzysztof Dzieciolowski supported by MITACS, DAESYS, Fin-ML and CUPFA.

Risk-Aware Bid Optimization for Online Display Advertisement

Date : November 4, 2022

This research focuses on the bid optimization problem in the real-time bidding setting for online display advertisements. We proposed risk-neutral and risk-averse models that maximize the expected profit for the advertiser by exploiting historical data to design an upfront bidding policy, mapping the type of advertisement opportunity to a bid price and accounting for the risk of violating the budget constraint during a given period of time. And, we present the numerical results from experiments, which demonstrate that our risk-averse method can effectively control the risk of overspending the budget while achieving a competitive level of profit compared with the risk-neutral model and a state-of-the-art data-driven risk-aware bidding approach.

By Rui FAN, a master's graduate in Data Science & Business Analytics at HEC Montreal and GERAD, and a current A.I. and Data Engineer at the Royal Bank of Canada. Her areas of research include optimization and machine learning. She is particularly interested in risk-aware optimization, integrated machine learning with operation research, and its applications in finance and e-commerce industries. Rui was also a recipient of Fin-ML CREATE grant supervised by Professor Érick Delage.

Foreseeing the Worst: Forecasting Electricity DART Spikes

Date : September 16, 2022

Statistical learning models are proposed for the prediction of the probability of a spike in the electricity DART (day-ahead minus real-time price) spread. Assessing the likelihood of DART spikes is of paramount importance for virtual bidders, among others. The model's performance is evaluated on historical data for the Long Island zone of the New York Independent System Operator (NYISO). A tailored feature set encompassing novel engineered features is designed. Such a set of features makes it possible to achieve excellent predictive performance and discriminatory power.

Results are shown to be robust to the choice of the predictive algorithm. Lastly, the benefits of forecasting the spikes are illustrated through a trading exercise, confirming that trading strategies employing the model predicted probabilities as a signal generate consistent profits.

By Remi GALARNEAU-VINCENT, a Ph.D. candidate in Financial Engineering at HEC Montréal with Professor Geneviève Gauthier. He discovered the electricity market during a research partnership with the firm Plant-E corp. He is a Machine Learning enthusiast.

Generate and predict multivariate time series

Date : March 11, 2022

Time series forecasting is to try to predict future realizations based on past observations. Those predictions can be precise or probabilistic depending on the model. Recent development in machine learning has accelerated the quantity of interesting models. The model selection is very much dependent of the use case and needs a thorough investigation. It is however harder to compare different models to measure their prediction quality, robustness, computational performance and data requirements. We will present a package written in Python to measure the performance of multiple models. For that purpose, we will focus on the simulation of time series to use as benchmarks.

By Francis HUOT-CHANTAL, a Ph.D. candidate in Applied Mathematics at the Université de Montréal with Professor Fabian Bastin. The subject of his PhD research is mostly high Frequency market making. He likes to pursue theoretical proof of convergence. However, application is the most important part of his thesis. For that purpose, he mostly uses the programming languages Julia and Python.

Valuation of syndicated loan adjusted for the presence of financial covenant

Date : February 11, 2022

We evaluate the impact of including safety covenant in syndicated loan agreements. We propose a stochastic dynamic game model of syndicated loan contract adjustments in the presence of the covenant. The model accounts for the lender’s right to punish or tolerate any breach of the covenant, and for the borrower’s flexibility in adjusting its investment and risk-taking strategy. Our numerical experiments show that, while a safety covenant improves the loan value in most states, it can have an adverse effect when bankruptcy risk becomes important. Additional investigation shows that the lender can optimally tolerate some technical default to prevent this adverse effect. We also find that the value of covenant decreases as the borrower’s creditworthiness improves.

By Tiguene NABASSAGA, a Ph.D. candidate in Financial Engineering at HEC Montréal with  Professor Michèle Breton. Tiguene is currently Manager at Ernst & Young Global Consulting Services.

Deep unsupervised Anomaly Detection in the derivatives market

Date : December 3, 2021

The recent rise in the interest of market participants towards financial derivatives directly lead to an augmentation in the quantity of high-frequency data processed continually by all exchanges. Regulators, who constantly monitor markets to unveil potential infractions, still perform their investigations manually with the help of deterministic rule-based algorithms. Thus, this notable growth in data generation poses a heightened risk of fraudulent orders going unnoticed.
I will present an ongoing research work that aims to solve the problem of algorithmic fraud detection in the derivatives market. This work, in collaboration with TMX, relies on the vast literature of anomaly detection to locate potential fraudulent patterns in the limit order books’ time series. Because of the complex nature of that data and the very low amount of real fraud cases, a new framework based on deep unsupervised learning algorithms specifically designed for high-frequency time series is proposed.

By Cédric POUTRÉ, a Ph.D. candidate in Financial Mathematics at the Université de Montréal with Professor Manuel Morales. He is interested in the diverse applications of Machine Learning in high-frequency trading and market microstructure.

An Input-Output HMM to describe states of energy price in NYISO Market

Date : November 26, 2021

The energy market in New York is centralized and managed by an independent system operator called NYISO. The price of energy on this market is determined by an auction mechanism. The Analysis of this price reveals that it is very volatile for several reasons (uncertain and inelastic demand, dynamic changes in the behavior of participants, etc.) that can be represented by models indexed by Markov chains.
In this work, an Input-Output HMM models is used to predict the price of energy and describe its hidden states. This model is well known for its ability to predict conditional price density. In order to be able to better describe the states (especially if they are numerous), a parsimonious parameterization is used.

By Ismael ASSANI, a Ph.D. candidate in Statistics at the Université de Montréal with Professor Maciej Augustyniak. He is interested in the modeling of Financial Series particularly with model indexed by Hidden Markov chains (discrete and continuous), the Hedging and Pricing of derivatives as well as the application of Machine Learning methods in Finance.

Using Reinforcement Learning to maximize Customer Profitability and CLV at Financial Institutions

Date : November 19, 2021

Customer Lifetime Value, CLV, is a popular measure to understand the future profitability of customers to allocate resources in more efficient ways to keep the company alive during difficult economic situations.
We use Machine Learning tools to predict the expected revenue from each customer during one year of his/her relationship with the institution as the CLV of the customer. The approach is implemented on two datasets from two international financial institutions. Different feature engineering techniques were applied to improve the prediction power of the model. We used two stage or three stage prediction models.
In the second phase, we train a reinforcement learning algorithm based on the history of marketing activities and the CLV as the state of customers to determine the optimum marketing action for customers in each state to maximize their profitability.

By Meisam SOLTANI. With a background in Electrical Engineering and Automotive Industry, Meisam moved to Data Science and Analytics world. He graduated from the University of Waterloo Management Sciences program in 2014. While pursuing his Ph.D. in Management Analytics at Queens University with Professor Mikhail Nediak, he has been involved in many Data Science and Analytics projects in Scotiabank.

Enhanced Rating Prediction with Text Data

Date : October 22, 2021

Over recent years, Natural Language processing (NLP) is widely used in many areas including finance and economics. Over the last decades, some influential papers have investigated whether negative words in the press can affect significantly affect individual stocks and aggregate market (Tetlock, 2007). They provide strong empirical evidence between their dictionary and 10K/10Q filing and return, trading volume, unexpected earning, and fraud.

The aim of this research is to fill gap and provide a framework that assess the credit rating of the firm more broadly using natural language processing from firm fillings. The proposed model can be used to assess the discriminating power of the identified variables that have been shown to be related to the rating.

By Ernest Tafolong, a Master's graduated in Data Science at HEC Montreal with Professor Erik Delage. Ernest is currently a Manager of the Financial Engineering department for an established alternative credit company in Canada where he leads all scoring & provisioning Machine Learning models development for autos and personal loans. His previous experience includes Market Risk Quantification at the National Bank of, he also served as Advisor–Financial Engineering at Desjardins where he designed and implemented quantitative models for investment strategies and risk Management. Besides, he provided quantitative advisory for Model Risk Management at leading French banks. He also graduated with a Master in Financial Engineering in 2011 at the Université Laval.

Natural Language Processing (NLP) for fraud detection in derivatives market

Date : June 17, 2021

With increasing activities in the financial derivatives market, exchange regulators are seeking to build smarter market surveillance systems to detect potential frauds. Current systems are often based on rules capturing known suspicious patterns reflected in the structured market data, while not having the ability to process information conveyed by unstructured textual data such as business news, which, however, can have crucial real-time impact on the market. Thus, it is of great interest to leverage NLP to extract analytics from textual data, so that the surveillance system could assess trading behaviors more accurately due to the added awareness of the business context.
In this talk, I will introduce a work in progress that develops a financial events analysis NLP pipeline that extracts important events from financial news then analyzes their characteristics, and a framework to leverage those analytics to help detect financial frauds, especially illegal insider trading, in the derivatives market.

By Pan LIU, a Ph.D. candidate in Data Science at HEC Montreal with Professor Gilles Caporossi. His research interest: developing Natural Language Processing, and Machine Learning models to solve problems in Finance.

Options pricing via Neural Stochastic Differential Equations and Martingale Pricing Theory

Date : May 28, 2021

This research investigates pricing financial options based on the traditional martingale theory of arbitrage pricing combined with neural SDEs. We treat neural SDEs as universal Itô process approximators. In this way we can lift all assumptions on the form of the underlying price process, and compute theoretical option prices numerically based on empirical data.
Along the way we propose a new variation to train neural SDEs by implementing the Wassterstein distance metric as a loss function. While similar to the GAN formulation, this technique does not need a discriminator and works will in the experiments, albeit limited in scope. Furthermore, it is conjectured that the error of the option price implied by the learnt model can be bounded by the very Wasserstein distance metric that was used to fit the empirical data.

By Timothy DeLISE, a Ph.D. candidate in Mathematics at the Université de Montréal with Professor Manuel Morales. His research interests are: Mathematics, Machine Learning, and Software Development.

Deep hedging methods for pricing and hedging financial derivatives

Date : May 14, 2021

I will present a reinforcement learning approach for the problem of optimally hedging financial derivatives. I will start by giving an overview of the hedging problem with classical approaches considered in the literature.
Afterwards, I will present the approach called deep hedging which I extensively studied in my PhD research thesis. This class of algorithms represents trading policies by neural networks. In other words, asset positions used for hedging are obtained as the output vectors of neural networks. Various numerical experimentations of this approach will be presented for pricing and hedging different financial derivatives.

By Alexandre CARBONNEAU had obtained a Ph.D. in Financial Mathematics at Concordia University with Professor Frédéric Godin. He has a Bachelor's degree in Mathematics with a specialization in Actuarial Science, as well as a Master's degree in Mathematical and Computational finance from the Université de Montréal.

Comparative analysis of time series prediction libraries

Date : May 7, 2021

Les différents modèles permettant l'analyse et la prédiction de séries chronologiques sous le langage de programmation Python sont listés et catégorisés. Ensuite, des données sur sept (7) contrats à terme différents sont obtenues et nettoyées afin d'être utilisées lors de l'analyse comparative. Une division en 20 échantillons d'entraînement et de test est réalisée dans le but d'évaluer l'erreur de prédiction. Une sélection restreinte de librairies adaptées au présent cas d'utilisation est effectuée. Le modèle ARIMA émerge comme étant le meilleur modèle en termes du MAE vis-à-vis les autres types de modèles pour la série S2.
Dans le but d'explorer plus en détails les modèles d'apprentissage profond, ceux-ci sont revisités. L'architecture du réseau de neurones récurrents à mémoire court et long terme (en anglais LSTM) est retenue comme étant la plus performante.
Enfin, ces résultats sont discutés puis des pistes d'amélioration sont présentées.

By Samuel TREMBLAY, Data Engineering Advisor at Mouvement Desjardins, where he assists many business sectors in the valorisation of their big data. He had a Bachelor’s degree in Aerospace Engineering and a Master in Business Intelligence from HEC Montréal with Professor Jean-Francois Plante and Professor Gregory Vial.

Bond price prediction with filter-based RNN and arbitrage-free regularization

Date : April 23, 2021

We review the class of HJM forward rate model and the dynamic term structure model in forecasting the yield curves and the prices of the coupon bonds by filter-based recurrent neural networks.
We propose model in the Kalman filter and the particle filter with the arbitrage-free restriction as regularization using the deep neural networks. We mainly build on LSTM to extract the useful information from historical time series of daily closed bond data and other type of networks to process the sequential errors for yield curve modeling and calibrating.
The purpose of the research is mainly to study the effect of arbitrage penalty on different forecasting horizons. We use data from U.S. Treasuries and coupon bonds of twelve corporate issuers. We provide analysis of the prediction error, error distribution, the average excess return, and results of out-of-sample test to show the efficiency and performance of our models.

By Xiang GAO had obtained a Ph.D. in Applied Mathematics at Concordia University with Professor Cody Hyndman. His research interests are: Financial mathematics and Machine Learning.

A WaveNet based model for Portfolio Management by using Deep Reinforcement Learning

Date : April 16, 2021

Portfolio management is a challenging task in its nature because of the abundance of factors that should be simultaneously considered: the investors' risk preferences and constraints, the investment environments and their limitations, and complicated features that affect the future price movements.
In this study, we propose a model based on actor-only reinforcement learning that takes all of these factors into account and provides a significant improvement in the portfolio performance compared to the most recent deep reinforcement learning models. In particular, we build on WaveNet, a model that was originally proposed for generating audio waveforms, to extract the useful information from historical time series of assets for portfolio management. Contrary to the similar previous models, we show that the performance of our approach is not affected by how the assets are permuted when fed into the model, and this could make the output of the model more reliable by practitioners.


Building a Recommender System using Graph Neural Networks

Date : February 12, 2021

Recommender systems are now a well-known technology used by many e-commerce platforms. Those platforms often have access to diverse sources of data, and traditional recommender systems might not accurately leverage all the available information.
This conference introduces the usage of Graph Neural Networks (GNNs) for personalized recommendation in a large-scale and heterogeneous data setting. We will observe how popular GNN models can be used for the task and what enhancements worked best for Decathlon, notably concerning how to handle evolving and seasonal customers’ preferences.

Hedging Basket Options with Deep Learning

Date : January 29, 2021

This conference introduces an effective model-free hedging strategy for a Basket Option using its component assets. By using Principal Components Analysis, the cost of trading can be reduced significantly and mitigate the effect of trading with non-liquid assets. Through numerical analysis, we will observe that an LSTM architecture for solving such a problem is very robust to compute an optimal solution. The model will also express the capacity of the neural network to fit an hedging strategy based on a custom risk function of an agent.