meta.py

"""
Big meta study, cross all the data with all the models

"""

from datasets import *
from metrics import *
from trend import TrendForecaster
from sgt import SGT
from micro import Micro
from tfp_sts import TfpSts
from prophet_model import ProphetModel
from fourex import FourierExtrap
from recency import RecencyForecaster
from diff_tree import *
from bandit import *
from pop import *
from greyk import *

import pandas as pd

from sktime.forecasting.naive import NaiveForecaster
from sktime.forecasting.theta import ThetaForecaster
from sktime.forecasting.ets import AutoETS
# from sktime.forecasting.trend import STLForecaster
from trend2 import STLForecaster
from sktime.forecasting.arima import AutoARIMA
from sktime.forecasting.exp_smoothing import ExponentialSmoothing
from sktime.forecasting.croston import Croston


results = []
datas = [load_daily_data(100), load_weekly_data(150), load_zero_bound_data(100), 
         load_btc_data()]
sps = [7, 52, 1, 7]
fh = np.arange(1, 11)
trains = [np.arange(0, 90), np.arange(0, 140), np.arange(0, 90), np.arange(0, 300)]
tests = [np.arange(90, 100), np.arange(140, 150), np.arange(90, 100), np.arange(300, 310)]
models = [NaiveForecaster, 
          TrendForecaster, 
          Croston, 
          ExponentialSmoothing, 
          AutoETS, 
          AutoARIMA, 
          ProphetModel, 
          SGT, 
          ThetaForecaster,
          Micro, 
          # TfpSts(), 
          FourierExtrap, 
          RecencyForecaster,
          DiffTree, # 12
          PoPForecaster,
          STLForecaster,
          Bandit,
          Greykite
          ]

for i, data in enumerate(datas):
  train = trains[i]
  test = tests[i]
  sp = sps[i]
  for j, model in enumerate(models):
    if j == 3: # exp smoothing
      model_inst = model('add')
    elif j == 8: # theta
      model_inst = model(deseasonalize=False) 
    elif j in [9, 16]: # micro or greykite
      model_inst = model(k=max(fh))
    elif j == 12: # difftree
      model_inst = model(fh=fh)
    else:
      model_inst = model()
    if hasattr(model_inst, 'sp'):
      model_inst.sp = sp
    if j == 14 and sp == 1: # stl
      model_inst.sp = 2
    model_inst.fit(data.y[train])
    pred = model_inst.predict(fh)
    # record mae and r2
    err = mae(data.y[test], np.array(pred))
    c = corr(data.y[test], np.array(pred))
    b = bias(data.y[test], pred)
    results.append((i, model_inst, err, c, b))
    print(i, model_inst, err, c, b)

df = pd.DataFrame(results)
df.to_csv('results.csv')