Amazon Forecast 与 StatsForecast 对比

!pip install statsforecast s3fs

import pandas as pd

Y_df_m5 = pd.read_parquet('https://m5-benchmarks.s3.amazonaws.com/data/train/target.parquet') 

Y_df_m5 = Y_df_m5.rename(columns={
    'item_id': 'unique_id', 
    'timestamp': 'ds', 
    'demand': 'y'
}) 

Y_df_m5.head()

from statsforecast import StatsForecast
from statsforecast.models import (
    AutoETS,
    DynamicOptimizedTheta,
    SeasonalNaive
)

# Create list of models
models = [
    AutoETS(season_length=7),
    DynamicOptimizedTheta(season_length=7),
]

# Instantiate StatsForecast class
sf = StatsForecast( 
    models=models,
    freq='D', 
    n_jobs=-1,
    fallback_model=SeasonalNaive(season_length=7)
)

/home/ubuntu/fede/statsforecast/statsforecast/core.py:21: TqdmExperimentalWarning: Using `tqdm.autonotebook.tqdm` in notebook mode. Use `tqdm.tqdm` instead to force console mode (e.g. in jupyter console)
  from tqdm.autonotebook import tqdm

from time import time

init = time()
forecasts_df = sf.forecast(df=Y_df_m5, h=28)
end = time()
print(f'Statsforecast time M5 {(end - init) / 60}')

Statsforecast time M5 14.274124479293823

forecasts_df['ThETS'] = forecasts_df[['DynamicOptimizedTheta', 'AutoETS']].clip(0).median(axis=1, numeric_only=True)
forecasts_df.to_parquet('s3://m5-benchmarks/forecasts/statsforecast-m5.parquet')

!pip install datasetsforecast

from datasetsforecast.m5 import M5, M5Evaluation

from datasetsforecast.m5 import M5, M5Evaluation
from statsforecast import StatsForecast

### Evaluator
def evaluate_forecasts(df, model, model_name):
    Y_hat = df.set_index('ds', append=True)[model].unstack()
    *_, S_df = M5.load('data')
    Y_hat = S_df.merge(Y_hat, how='left', on=['unique_id'])
    eval_ = M5Evaluation.evaluate(y_hat=Y_hat, directory='./data')
    eval_ = eval_.rename(columns={'wrmsse': f'{model_name}_{model}_wrmsse'})
    return eval_

### Read Forecasts
statsforecasts_df = pd.read_parquet('s3://m5-benchmarks/forecasts/statsforecast-m5.parquet')
amazonforecasts_df = pd.read_parquet('s3://m5-benchmarks/forecasts/amazonforecast-m5.parquet')

### Amazon Forecast wrangling
amazonforecasts_df = amazonforecasts_df.rename(columns={'item_id': 'unique_id', 'date': 'ds'})
# amazon forecast returns the unique_id column in lower case
# we need to transform it to upper case to ensure proper merging
amazonforecasts_df['unique_id'] = amazonforecasts_df['unique_id'].str.upper()
amazonforecasts_df = amazonforecasts_df.set_index('unique_id')
# parse datestamp
amazonforecasts_df['ds'] = pd.to_datetime(amazonforecasts_df['ds']).dt.tz_localize(None)

### Evaluate performances
m5_eval_df = pd.concat([
    evaluate_forecasts(statsforecasts_df, 'ThETS', 'StatsForecast'),
    evaluate_forecasts(statsforecasts_df, 'AutoETS', 'StatsForecast'),
    evaluate_forecasts(statsforecasts_df, 'DynamicOptimizedTheta', 'StatsForecast'),
    evaluate_forecasts(amazonforecasts_df, 'p50', 'AmazonForecast'),
], axis=1)
m5_eval_df.T