Deep Learning for finance has always been applied through a wealth of techniques and network architectures to try to predict the evolution of financial instruments and specifically stock markets. A diversity of new sources such as tweets have also being used to gain superior predictive power. We suggest in this project for the course CS230-Deep Learning to apply feature knowledge used in factor investing to evaluate if a Q-learning network can better identify investments opportunities in the equity markets.

Several methods such has for example the Kalman filter have been used at identifying and correcting forecasts as the market reveals more information as time updates.

• Factor investing pioneered by FamaFrench has set the base of Quantitative Investment with the 3 factors: size of firms, book-to-market values and excess return on the market.
• KahnGinold introduced the concept of the selection of factors with the Information Ratio (IR) that consists of a measure of the value added per unit of risk of each independent factors (Information Coefficient (IC)) times how often that factor can be put in use in a trading setting (Breadth).

Thus, historically, factor investing has focused on identifying factors returns and their respective correlations to generate superior returns. Methods have ranged from ordering factor Z-scoring, moving-average z-score as presented in QuantEquity.

To investigate the methods of Deep Learning in a context of identifying factors and their Information Coefficient to implement factor investing, DRLinPort and FactorInRL point in interesting directions in using Deep Reinforcement Learning. DRLinPortcompares different type of Neural Networks (LSTM, CNN, RNN ) to build optimal Portfolio through policy functions. FactorInRL is more interesting in focusing in factor investment by building a genetic algorithm to select the factors that will build the portfolio.
DRLinPort notably mention the fact that the genetic algorithm is superior in that it will not be subject of several local minimum compared to gradient descent methods.

This project plans to investigate/build the model below that identifies IC factors though a policy function by forecasting rewards with simulation of expected returns.

Model Description:

• Dimensionality of the features W_nm,nz in the DNN
  • n_m with the number of stocks from the S&P500 selected,
  • n_z the number of alpha-factors under consideration
• State: represent the vector of the positions in stock
• Agent: the agent is simply be the change of positions in the n_m stocks that maximises the returns r_k over the horizon k. That would be based on a Q-table using states of each of the stocks and the actions for each stock

KahnGinold Information Coefficient (IC) and alpha formula:

Structure:
	Readme.md
	/Data-snapshot
		/Commodities
		/Commodities Static
		/Description
		/Economic
		/Economic Poll
		/Economic Static
		/Futures
		/Futures Static
		/FX
		/FX Static
		/Index
		/Index Static
		/Interest Rates
		/Interest Rates Static
		/Options Index
		/Options Index Static
		/US Stocks
		/US Stocks Static
	/Data-processed
	/Data-factor
	/Output
    /project_report
        CS230_project_report.pdf

Data description:

• "Data-snapshot" ("Data" can be obtained through downloading at DataStream Refinitiv) contains the raw time-series of market informations (Stocks, Commodities, Economic information, Futures, Currencies(FX), Stock and Credit Indices, Interest Rates, Options on Stock Indices). "Static" extension contains the information describing the nature of the financial contracts, i.e. for a stock options: the expiry, the underlying stock, the name of the contract: Put or Call, etc.

• "Data-processed" contains the information of raw time-series grouped along the time axis $t$ split in 2 arrays: (Not uploaded - Currenty under development)

  • data_stocks collects from "Data-snapshot" the time-series of the values of the stocks to help compute the $Q(s,a)$ function by simulating the value of the agent composition of the positions searched.
  • mrkt_info collects from "Data-snapshot" the time-series of the market conditions such as the implied volatility, the momentum factors, positions of currencies, oil prices etc. This data is usefull in constructing the Information Coefficient (IC) factors that will feed the Deep Neural Network in the model. This is the dataset used for the Back-testing in the model.
  • mrkt_info collects from "Data-snapshot" the time-series of the market conditions such as the implied volatility, the momentum factors, positions of currencies, oil prices etc. This data is usefull in constructing the Information Coefficient (IC) factors that will feed the Deep Neural Network in the model.

• "Data-factor" contains the time-series of factors build from mrkt_info that will feed the DNN in the model.

• "Output" contains the time-series of the positions taken by the Q(s,a) function.

@article{FamaFrench,
  title   = "Common risk factors in the returns on stocks and bonds",
  journal = "Journal
of Financial Economics" ,
    volume  = "33",
  pages   = "3--56 ",
  year    = "1993",
  author  = "Fama, E. F. and K. R. French"
}

@article{KahnGinold,
  title   = "Active Portfolio Management : A quantitative approach for producing superior returns and selecting superior money managers",
  journal = "McGraw-Hill Education" ,
  pages   = "-- ",
  year    = "1999,
  author  = "R. Grinold and R. Kahn"
}

@book{QuantEquity,
  title   = "Quantitative equity portfolio management: Modern techniques and applications",
  publisher = "Chapman and Hall/CRC Financial Mathematics Series." ,
  pages   = "462",
  year    = "2007,
  author  = "E. E. Qian, R. H. Hua, E. H. Sorensen"
}

@article{DRLinPort,
  title   = "A deep reinforcement learning framework for the financial portfolio management problem",
  journal = "Xi’an Jiaotong-Liverpool University",
  pages   = "-- ",
  year    = "Jul 2017",
  doi     = "–doi:arXiv:1706.10059v2.",
  author  = "Z. Jiang, D. Xu, J. Liang"
}

@article{FactorInRL,
  title   = "Factor Selection with Deep Reinforcement Learning for Financial Forecasting",
  journal = "University of Chicago, McKinley Capital Management",
  pages   = "-- ",
  year    = "Oct 2018",
  doi     = " https://ssrn.com/abstract=3128678",
  author  = "Ziwei Wang and Nelson Leung"
}

Released under MIT License. See [LICENSE][license] for details.