Machine Learning methods in Quantitative Finance

Learning outcomes
After the completition of this course students will be able to:
  • Measure quantitative strategy profits in the light of Fama-French risk factor model.
  • Implement forecasting models using multiple Machine Learning tools: Gaussian Process, Rule Learning and Artificial Neural Networks.
  • Backtest portfolio allocation strategies employing tools from Online Machine Learning theory (Universal Portfolio)
  • Employ forecasting models based on the Sentiment Analysis (Natural Language Processing) of a rich proprietary dataset of Twitter posts. The dataset consist of all tweets of Down Jones stocks since Twitter inception.
The summer school will have lectures, practicals, and self-reflection sessions, which might culminate in additional homework.
The practicals make use of real time series (daily and high-frequency) and of our Twitter datasets (proprietary).
  • Day 1: Market Phenomenology and Stylized Facts
    • Common risk factors of stock returns
    • Exotic risk factors (currency carry, volatility premia, interest rate curve premia and so on)
    • Fama-French Model: The difference between portfolio $\alpha$ and $\beta$
    • Fitting the Fama-French Model for multiple stocks
    • Is it easy to generate true $\alpha$ investment strategies?
    • Do you agree with the Fama-French hypothesis regarding $\alpha$? How to test if stock prices follow Random Walks?
  • Day 2: Classical Time Series Methods
    • Stationary and non-stationary time series: $ARMA$ and $ARIMA$ models
    • The high predicability of market volatility: $GARCH$ models
    • Time series Cointegration: Quant's bread and butter
    • Testing Cointegration of stocks in the same sector/industry
    • Backtesting a simple trading strategy based on Autoregressive models
    • Cross-validating the praticals outcomes
  • Day 3: Introducing non-linearities and capturing Seasonal patterns
    • Exploring Equation Discovery frameworks
    • Gaussian Processes: A powerful tool
    • Forecasting Economic Indicators: Discovering equations
    • Predicting stock prices using Gaussian Process
    • How to extend Gaussian processes to more than one regressor
  • Day 4: News as an exogenous variable
    • Reviewing the Financial literature: Market Over/Underreaction to news
    • Rational Expectations and Behavioral Finance explanations of stylized facts
    • Sentiment Analsys: A Machine Learning task (Dictionary-based and Supervised methods)
    • Build a Sentiment classifier using our proprietary Dow Jones Tweets Dataset (a dataset with more than 1 million instances)
    • How to scale Sentiment Analysis to unlabelled data?
  • Day 5 (half-day): Natural Language Processing and the Anatomy of Volkswagen Sacandal
    • Analyse the VW scandal using our proprietary database of Tweets (all Tweets with keywords related to the VW scandal) and our intraday (1-minute) stock price series
  • Day 6: Markets and regime switching models
    • Hidden Markov Models (HMM): The occasionally dishonest casino example
    • A market with two markov regimes: AR(1) and Random Walk
    • Fitting HMM to stock price data
    • How persistent are the market regimes?
    • How to implement a control system feature to a backtesting aiming to avoid drastic losses during the switching phases
  • Day 7: Online Machine Learnig (OML) and Portfolio Theory
    • General OML problems review
    • Markowitz Portfolio Optimization (quadratic convex optimization)
    • Dirichlet distribution and random portfolio weights
    • Bringing OML to the Portfolio Selection problem: Universal Portfolio and the Best Constant Rebalanced Portfolio (BCRP)
    • Backtesting the Universal Portfolio updating algorithm
  • Day 8: Supervised Learning -- Feature Engineering and Learning representations
    • Decision Tree Learning and applications (Credit Scoring and stocks next period return classification)
    • Ensemble Methods: Random Forest
    • Deep Learning and Convolutional Networks: Treating the price graph as an image and classifiyng next period returns
    • Backtesting next day and next five days periods returns using a "toy model" with features depending on: price convolutions, price mean-reverting indicators and Open/High/Low/Close prices
    • In terms of portfolio performance, is representations learning as good as hand-crafted fetures? What additional features could we add?
  • Day 9: Hackathon
    • Putting past prices and Sentiment analysis indicators together. Competing for the most profitable model!