TradeML is a full-stack web app with a full-fledged Rescript React frontend and OCaml Dream backend. With the latest historical data of Bitcoin prices and related financial signals (e.g. volumn), we built a deep neural time series forecasting model for Bitcoin price using stacked stateless Long-Short Term Memory (LSTM)[1]. On top of this, we created a simulation game where the user gets to hypothetically trade Bitcoin from a wallet, using the predicted near-future price to facilitite the decision and trade with real-time price. It also comes with a linechart visualization of recent transactions, e.g., changes in total assets and Bitcoin prices.
[1] - The choice of this architecture is based our literature review of the following paper: https://arxiv.org/abs/2004.10240
Repo: https://github.com/robertzhidealx/btc-game-monorepo
Slides: https://docs.google.com/presentation/d/1vWWM401cLiOVfWBdSuIutFDIxt2f0hfR-gOhiIhl52o/edit#slide=id.p
Frontend deployment: https://trade-ml.vercel.app. (Analystics page down due to an issue with a third-party JavaScript library (ReCharts), but works just fine locally. See the App section for details.)
Team members and responsibilities:
- Jiaxuan Zhang (jzhan239): data retrieval and shaping functions, database module (
DBmodule), game module (Gamemodule), the backend server (app.ml- Dream server and related functions), and the frontend web app (React, written in Rescript) - Chuheng Hu (chu29): forecasting model (
./server/forecasting/modelfiles (data processing & training), Forecast module (loading & inference)), and the backend Visualization module and related Dream server function
This monorepo contains the code for both the frontend and backend of our game. app contains code pertaining to the Rescript frontend, and server contains code pertaining to the OCaml backend.
.
├── app
│ ├── public
│ │ └── svg
│ └── src
│ ├── components
│ ├── pages
│ └── utils
├── assets
└── server
├── _coverage
│ └── src
├── forecasting
│ ├── data
│ └── model
├── src
└── test
server/src contains business functions and the Dream server
server/forecasting/data contains the data set we use for training our time series forecasting model and backtesting.
server/forecasting/model contains the code for training the model which is implemented in python.
The code coverage for the backend is at 100% (for testable code). For instructions on running code coverage, please refer to the Server Run section.
Frontend:
.res files ~1250 LoC
Backend:
.ml files ~850 LoC
.mli files ~180 LoC
Model training:
Python ~130 LoC
Total ~2400 LoC
Build the entire project at the root by running dune clean and dune build. For details setting up server and app, refer to the following documentation.
First, run a dune build under the server directory. Install any missing libraries by following their docs listed below to install using opam install x.
Note that for the
ocaml-cohttplibrary specifically, please make suretlsis locally installed viaopam install tls. This is not well-documented by the offical docs.
Install PostgreSQL-OCaml via
opam install postgresql
Caution: The ocaml-torch library is NOT compatible with Apple M1 machines, leading to errors on
dune build. In this case, comment outmodule Forecast ...in bothlib.mlandlib.mli, and comment out calls tomodule Forecastfunctions inapp.mlto build other parts of the project.
Listed below are the official docs of the libraries used.
You can install them using pip/pip3.
First run a fresh dune clean and dune build in the server directory. Then configure the database as follows.
Install Postgresql on your machine via brew install postgresql (MacOS).
Prereq: Make sure there is a Superuser named
postgresin your local Postgresql server. A guaranteed solution is tocreatedb db->psql db->CREATE USER postgres SUPERUSER;-> exitpsql->dropdb dband proceed with the following steps. (See this post for reference.)
Start the local database server (MacOS) by running
brew services start postgresql
and create a database named testdb by running
createdb -h localhost -p 5432 -U postgres -O postgres testdb
to start the local postgresql database server, and then run
dune exec ./src/app.exeto execute the app compiled from app.ml. Supporting functions are in lib.ml as specified by lib.mli.
Please note that you may have to connect to a personal hotspot or a VPN (and especially not the
hopkinsnetwork) for the GET request to Binance endpoints to go through, and ultimately for the server to work correctly. As crypto adoption progresses, the regulatory actions of the crypto industry and exchanges are making progress in the U.S. recently.
To run tests and coverage, cd into ./server and run dune test. After this, run bisect-ppx-report html and then open ./_coverage/index.html to see results. The coverage is now at 100%.
Cd into the app directory. Make sure you have Node.js and NPM installed locally. The frontend app is written in Rescript and uses rescript-react. I recommend installing the rescript-vscode VSCode extension for syntax highlighting and intellisense.
Make sure that the Server code is running via the aforementioned steps.
First run npm install to set up the dependencies. Then, you need to open two terminal tabs, one is where you run npm run start to start the Rescript compiler in watch mode, the other is where you run npm run server to start the local development server. This is all it takes to start the web app.
I had Vercel wired up such that we would always have the latest production build deployed at https://trade-ml.vercel.app/, so feel free to try our app out there. Please note that the visualization (analytics) part was broken in production due to the underlying library not compiling correctly, but it works fine locally.
Currently, the frontend web app is looking like this:
The main landing dashboard page:
The Bitcoin price forecaster:
The analytics (visualization) page:
