![](https://crypto4nerd.com/wp-content/uploads/2023/02/0e4ZqK7IDADD8KLwi.jpeg)
Deep Learning Project Proposal — Stanford CS230 and deeplearning.ai
Photo by Luca Bravo on Unsplash
Online courses in machine learning
I have been studying deep learning with Stanford’s CS230 videos and the detailed material on deeplearning.ai. Andrew Ng is a highly methodical professor, and while I went in wanting just a “wam bam, thank you m’am” approach to machine learning, I’ve found myself better off for knowing the math behind the algorithms (this is evident when I have no idea why something isn’t working the way I want it to, and I can dive into the code, do some linear algebra and correctly find my way out of the bug). There is a class project that needs to be completed by the end of the course, and I’m going to apply it to algorithmic trading. That is, using some of my day trading strategies, and seeing if I can train the bot to win the “game” using reinforcement learning using profit as points. There have been several attempts with the projects of other students over the last few years, with mixed results.
Project Proposal
Category — Reinforcement Learning I’ll be investigating whether or not I’ll be able to ‘gamify’ day trading in FOREX and in doing so, teach a neural network to optimize entry and exit times to maximize profitability. This is interesting because old market wisdom says that it’s easy to get in the market, but once you’re in, it’s hard to get out with profit. I’d like to see if I can identify entry and exit pairs that will result in a net profit. Challenges — Some of the challenges of the project are simple. It’s hard to predict the market. It’s been shown time and time again that it barely works in the long run, especially when you have different “general conditions” in the market, referring to the fact that a market can be generally bullish, generally bearish, or trending (sideways). I need to find a type of neural network architecture that can get into a market position, and decide to get out if conditions seem like they are against the position, vs hang on if there might be a resulting profit. I also need to consider the spread or commission inherent in trading, so that the profit must exceed the cost of making the trade and holding onto the trade. Additionally I need to find an algorithmic framework that allows me to control it with the neural network I plan to build. Dataset — I’ll be utilizing historical data for the past 5–10 years for GBP/USD and USD/MXN pairs. “Cable”, another name for GBP/USD, has high volatility and high volume, while “Dollar-Peso” being the 7th most traded pair, is a little less “up and down” and will provide a more stable pair to train on. I want to limit my dataset to the past 5 years if possible to account for the inclusion of mainstream trading algorithms, as the data from before then is much less volatile and exhibits different properties. I’ll be looking at the minutely charts to get the highest resolution, but may switch to the hourly chart if the strategy calls for it. This data will be sourced from oanda.com or forex.com depending on the broker that ends up being used. The reason being is that FOREX data is not centralized, and therefore each broker has their own historical values. An algorithm trained on one service will simply not work on another service. Proposed method or Algorithm — I propose to research into reinforcement learning algorithms but also to take a look at recurrent neural networks, this is because I may need to apply sequence information in looking at “setups”, knowing that any particular moment is good to get in depending on the previous data in the chart. Reading list — TBD Evaluation of Results — Results will be scored on percent profit, with more dollars being better than fewer. It will be considered a success if the profit is shown to be greater than a buy and hold strategy over each period the algorithm is applied to. Stay tuned for updates.