This project aims to predict the future price of mutual fund via GluonTS package.

• Refactor the folder structure:
  • Understand the purpose of each file
  • Move Similar files to the same folder
    • backtest folder contains all backtesting codes
    • [-] util folder contains files or tools that is used by many other codes
    • reference folder contains the notebooks and experiments
  • XXX: Small Refactoring:
    • Move fund_price_loader import out of backtest/ and import it in run_backtest.py and pass it to init of backtest/ from there.
    • Move NAV_DIR to config.py
    • Change billiard Pool to Ray ActorPool (more stable and useful!)
  • FIXME: Extension Enhancement:
    • Allow time-dependent factors in backtest/ to be generalized (Because the frequency of tick may increase if binance api is adopted).
• Price Prediction of Single Fund
  • Load fund price csv file into Gluonts time series object.
  • Process the time series so that the missing prices of holiday can be interpolate.
  • Connect the time series to the Gluonts Model.
  • Plot the predicted trends (probalistically) from the trained model.
• Enhance Evaluation:
  • Allow splitting of training, validation, testing time series.
  • Allow evaluation of prediction using RMSE on testing data.
  • Allow Backtesting (using Off-the-shielf module of Gluonts).
    • Parallelize BackTesting
      • split_date generator
      • split dataset -> prediction -> evaluation
    • [-] Try sharing of NAV Table Between Process
• Adapt the Model in backtesting to the Deep Trainable Models
  • using pytorch version models with pytorchts package: https://github.com/zalandoresearch/pytorch-ts
  • Adapt to all examples in pytorch-ts
    • Implicit Quantile Network
    • Multivariate-Flow # Next-Up
    • Time-Grad
• Refactor the current architechture such that adapting to Multi-Variate mode can be easier to follow.
  • Seperate nav splitting methods from fund_price_loader.py to nav_splitter.py
  • Refactor so that replication between backtesting and multi_variate_backtesting can be reduced.
    • A abstract BackTestBase object
    • A basic BackTest object
    • A multi-variate BackTest object
    • An object allow both single&multi-variate models
• Enhance the evaluation
  • adapt the evaluation scheme to more metrices: check implicit_quantile_network.py Line.52-67) (TODO: now-fbprophet is not working)
  • Allow comparison of different models in a single plot
• Consider MultiVariate Mode for Single Fund Prediction:
  • Find data object in gluonts for storing multiple time series (check multivariate_dataset_examples.py)
  • Organize of nav curves of multiple funds into the multi-timeseries objects offered by gluonts.
    • Read nav curves into multiple SharableListDataset
    • Using Spliter to obtain multiple train, test SharablesListDataset(s)
    • [build] a SharableMultiVariateDataset which allow storing of multiple sharable target arrays and allow convertion to grouped_list_dataset (check multivariate_dataset_examples.py for progamming the convertion).
    • Before convert those train, test to local ListDataset(s), merge them into SharableMultiVariateDataset
    • Convert the multivariate_dataset into local grouped dataset using train_grouper and test_grouper
  • Adapt to Multi-Variate Deep Model (see examples of pytorch-ts) and incorporate it into the repo.
    • Allow evaluation of Multiple Time Series (see plot and MultivariateEvaluator in Time-Grad-Electricity)
  • Create different technical curves for each fund
    • Earning of fund in a time period: e.g., (nav tomorrow - nav today) / nav today. (parameter: time_period)
    • Standard deviation of earning in a time periods. (parameter: earning_time_period, std_time_period)
    • Original NAV curve
• Price Prediction of Multiple Funds
  • Load multiple funds and convert to multiple time series
  • Parallel loading and processing of multiple time series
  • Consider multivariate time series Model
  • Plot the predicted trends
• Consider Binance Dataset
• Adoption of Real-time Binance Price
• Allow parameter tuning for each estimator (with ray.tune)

python3 -m virtualenv env
source env/bin/activate
sudo pip install --upgrade pip
pip install -r requirements.txt

source env/bin/activate

python backtesting.py