two_sigma_financial_modelling.py 文件源码

def portfolio_timestamp_period_with_most_highly_corr_assets(self, df):
        # A first approximation to model portfolio returns:
        # i) Find assets that correlates with y, where correlation is higher than a threshold value
        # ii) Include only above assets and find maximum timestamp period with most assets
        # iii) Transform target value y to be cumulative mean of y in order to obtain monotonic behaviour
        # iv) Train model to predict transformed target value with the selected most correlated assets in selected
        # timestamp interval
        # v) Run model on test data and apply inverse transform to get target value y.

        # From plot it looks like a lot of assets are bought and sold at first and last timestamp.
        # We should of course primarily select assets based on how much they are correlated with y

        correlation_coeffecients = self.correlation_coeffecients
        names_of_assets = correlation_coeffecients.loc[correlation_coeffecients.index != 'y'].sort_values(
            ascending=False).head(self.number_of_assets_in_portfolio).index
        # Todo: make a check if any intermediate sales assets are among the most corr with y
        return df.loc[:, names_of_assets]