Add residuals checks

[capelastegui]

Our prediction intervals sometimes have a trend, suggesting that the residuals for that model have a non-zero null. This points to a sub-optimal fit - we should check for it and fix it.

[capelastegui]

A first step would be to generate a residuals dataframe including all sources and models. Could would be based on the following:

l_cols = ['date', 'source']

df_actuals_unfiltered = df_forecast.loc[df_forecast.is_actuals &
                                        ~df_forecast.y.isnull()]
df_actuals = (df_actuals_unfiltered[['date', 'y']]
              .rename({'y': 'actuals'}, axis=1)
              )
# Compute residuals for filtered actuals
df_residuals_tmp = df_forecast.loc[
    ~df_forecast.is_actuals & ~df_forecast.y.pipe(
        pd.isnull)][l_cols + ['model', 'y']]
df_residuals = df_residuals_tmp.merge(df_actuals, how='inner')
df_residuals['res'] = df_residuals['actuals'] - df_residuals['y']