alphamodel

What is it meant for?

alphamodel is an alpha development tool meant to package data fetching, model training and model prediction. The base example models are designed to:

fetch historical data from Quandl or csvs
generate predictions (basic EWMA or HMM)
estimate a covariance matrix based on direct estimation or Fama-French factor models.

The base model outputs are standardized as inputs to the cvxportfolio library which can be used for portfolio optimization and back testing.

Can it do anything special?

Funny you should ask, yes! A custom application of alphamodel is its use to tell investors when to invest in their views and when to hold off. To achieve this, it uses Black Litterman return and risk estimates where:

the user can provide investment views in a linear combination based format and
the model automatically incorporates them together with an EWMA or HMM based confidence level a new set of output return and risk estimates.

Using this new set of estimates leads to a portfolio with the views incorporated (proportional with how much the model thinks they're likely to be active at that time).

Config

The configuration follows a simple yml format but can also be provided directly as a nested dictionary. See alphamodel/examples/ for 2 sample yml files.

alpha:
  name: rebalance_sim
  universe:
    path: '../data/SP100_2010.csv'
    ticker_col: Symbol
    risk_free_symbol: USDOLLAR
  drop_threshold: 0.5
  data:
    name: eod_returns
    source: quandl
    table: EOD
    api_key: 6XyApK2BBj_MraQg2TMD
  model:
    start_date: '20100102'
    end_date: '20171231'
    data_dir: '../data/'
    halflife: 4
    horizon: 1
    min_periods: 4
    returns:
      sampling_freq: weekly
    covariance:
      method: FF5
      sampling_freq: daily
      update: monthly

Examples

Please review the alphamodel/examples sub-folders for:

Jupyter notebooks with sample simulations and charts similar to the paper
Python scripts to rerun the full efficient frontier simulations

Remember to cite our paper

If using this library please cite the upcoming paper: