btc_ridge.py

# implements a rolling ridge regression of length window

window = 250 # train window

dt_list = []
pred_list = []
up1sd =[]
down1sd = []
rsq = []

pred_df = pd.DataFrame()
#xvars_df_zsc = (xvars_df_m - xvars_df_m.rolling(60).mean()) / xvars_df_m.rolling(60).std()
#reg_dat = pd.concat([xvars_df_zsc,y],axis=1).dropna()
beta_df = pd.DataFrame()

res = Ridge(alpha=0.1,normalize=True)

inputs = pd.DataFrame()

for i in range(window + 1,len(reg_dat.index)):
    df_train = reg_dat.iloc[(i-window-1):(i)].dropna()
    if(len(df_train) > 0):
        df_test = reg_dat.iloc[[i]]
        if(len(df_test.T.dropna()) > 0):
            dt_list = dt_list + [reg_dat.index[i]]
            x_train = df_train.drop(columns='Y_VAR')
            res.fit(x_train,df_train['Y_VAR'])
                        
            resid = df_train['Y_VAR'] - res.predict(x_train)
            res_sd = resid.std()
            roll_beta = pd.DataFrame(res.coef_).T

            roll_beta.index = [reg_dat.index[i]]
            beta_df = pd.concat([beta_df,roll_beta],axis=0)
            
            pred = res.predict(df_test.drop(columns='Y_VAR'))            
            inputs = inputs.append(df_test.drop(columns='Y_VAR'))
                        
            pred_list = pred_list + [pred]
            up1sd = up1sd + [pred+res_sd]
            down1sd = down1sd + [pred-res_sd]
            #rsq = rsq + [res.rsquared]
            
pred_df = pd.DataFrame(pred_list,index=dt_list,columns=['FORECAST'])
up_df = pd.DataFrame(up1sd,index=dt_list,columns=['UP_1SD'])
down_df = pd.DataFrame(down1sd,index=dt_list,columns=['DOWN_1SD'])

#rsq_df = pd.DataFrame(rsq,index=dt_list,columns=['RSQ'])
pred_df = pd.concat([pred_df,up_df,down_df],axis=1)
reg_pred_df = pred_df.copy()