!pip install statsforecast mlforecast neuralforecast datasetforecast s3fs pyarrow

import pandas as pd

# Load the training target dataset from the provided URL
Y_df = pd.read_parquet('https://m5-benchmarks.s3.amazonaws.com/data/train/target.parquet')

# Rename columns to match the Nixtlaverse's expectations
# The 'item_id' becomes 'unique_id' representing the unique identifier of the time series
# The 'timestamp' becomes 'ds' representing the time stamp of the data points
# The 'demand' becomes 'y' representing the target variable we want to forecast
Y_df = Y_df.rename(columns={
    'item_id': 'unique_id', 
    'timestamp': 'ds', 
    'demand': 'y'
})

# Convert the 'ds' column to datetime format to ensure proper handling of date-related operations in subsequent steps
Y_df['ds'] = pd.to_datetime(Y_df['ds'])

Y_df = Y_df.query('unique_id.str.startswith("FOODS_3")').reset_index(drop=True)

Y_df['unique_id'] = Y_df['unique_id'].astype(str)

from statsforecast import StatsForecast

/home/ubuntu/statsforecast/statsforecast/core.py:23: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html
  from tqdm.autonotebook import tqdm

# Feature: plot random series for EDA
StatsForecast.plot(Y_df)

# Feature: plot groups of series for EDA
StatsForecast.plot(Y_df, unique_ids=["FOODS_3_432_TX_2"])

# Feature: plot groups of series for EDA
StatsForecast.plot(Y_df, unique_ids=["FOODS_3_432_TX_2"], engine ='matplotlib')

# Import necessary models from the statsforecast library
from statsforecast.models import (
    # SeasonalNaive: A model that uses the previous season's data as the forecast
    SeasonalNaive,
    # Naive: A simple model that uses the last observed value as the forecast
    Naive,
    # HistoricAverage: This model uses the average of all historical data as the forecast
    HistoricAverage,
    # CrostonOptimized: A model specifically designed for intermittent demand forecasting
    CrostonOptimized,
    # ADIDA: Adaptive combination of Intermittent Demand Approaches, a model designed for intermittent demand
    ADIDA,
    # IMAPA: Intermittent Multiplicative AutoRegressive Average, a model for intermittent series that incorporates autocorrelation
    IMAPA,
    # AutoETS: Automated Exponential Smoothing model that automatically selects the best Exponential Smoothing model based on AIC
    AutoETS
)

horizon = 28
models = [
    SeasonalNaive(season_length=7),
    Naive(),
    HistoricAverage(),
    CrostonOptimized(),
    ADIDA(),
    IMAPA(),
    AutoETS(season_length=7)
]

# Instantiate the StatsForecast class
sf = StatsForecast(
    models=models,  # A list of models to be used for forecasting
    freq='D',  # The frequency of the time series data (in this case, 'D' stands for daily frequency)
    n_jobs=-1,  # The number of CPU cores to use for parallel execution (-1 means use all available cores)
)

from time import time

# Get the current time before forecasting starts, this will be used to measure the execution time
init = time()

# Call the forecast method of the StatsForecast instance to predict the next 28 days (h=28) 
# Level is set to [90], which means that it will compute the 90% prediction interval
fcst_df = sf.forecast(df=Y_df, h=28, level=[90])

# Get the current time after the forecasting ends
end = time()

# Calculate and print the total time taken for the forecasting in minutes
print(f'Forecast Minutes: {(end - init) / 60}')

Forecast Minutes: 2.270755163828532

fcst_df.head()

from mlforecast import MLForecast
from mlforecast.target_transforms import Differences
from mlforecast.utils import PredictionIntervals
from window_ops.expanding import expanding_mean

!pip install lightgbm xgboost

# Import the necessary models from various libraries

# LGBMRegressor: A gradient boosting framework that uses tree-based learning algorithms from the LightGBM library
from lightgbm import LGBMRegressor

# XGBRegressor: A gradient boosting regressor model from the XGBoost library
from xgboost import XGBRegressor

# LinearRegression: A simple linear regression model from the scikit-learn library
from sklearn.linear_model import LinearRegression

# Instantiate the MLForecast object
mlf = MLForecast(
    models=[LGBMRegressor(), XGBRegressor(), LinearRegression()],  # List of models for forecasting: LightGBM, XGBoost and Linear Regression
    freq='D',  # Frequency of the data - 'D' for daily frequency
    lags=list(range(1, 7)),  # Specific lags to use as regressors: 1 to 6 days
    lag_transforms = {
        1:  [expanding_mean],  # Apply expanding mean transformation to the lag of 1 day
    },
    date_features=['year', 'month', 'day', 'dayofweek', 'quarter', 'week'],  # Date features to use as regressors
)

# Start the timer to calculate the time taken for fitting the models
init = time()

# Fit the MLForecast models to the data, with prediction intervals set using a window size of 28 days
mlf.fit(Y_df, prediction_intervals=PredictionIntervals(window_size=28))

# Calculate the end time after fitting the models
end = time()

# Print the time taken to fit the MLForecast models, in minutes
print(f'MLForecast Minutes: {(end - init) / 60}')

MLForecast Minutes: 2.2809854547182717

fcst_mlf_df = mlf.predict(28, level=[90])

fcst_mlf_df.head()

# Read the results from Colab
fcst_nf_df = pd.read_parquet('https://m5-benchmarks.s3.amazonaws.com/data/forecast-nf.parquet')

fcst_nf_df.head()

# Merge the forecasts from StatsForecast and NeuralForecast
fcst_df = fcst_df.merge(fcst_nf_df, how='left', on=['unique_id', 'ds'])

# Merge the forecasts from MLForecast into the combined forecast dataframe
fcst_df = fcst_df.merge(fcst_mlf_df, how='left', on=['unique_id', 'ds'])

fcst_df.head()

sf.plot(Y_df, fcst_df, max_insample_length=28 * 3)

sf.plot(Y_df, fcst_df, max_insample_length=28 * 3, 
        models=['CrostonOptimized', 'AutoNHITS', 'SeasonalNaive', 'LGBMRegressor'])

init = time()
cv_df = sf.cross_validation(df=Y_df, h=horizon, n_windows=3, step_size=horizon, level=[90])
end = time()
print(f'CV Minutes: {(end - init) / 60}')

/home/ubuntu/statsforecast/statsforecast/ets.py:1041: RuntimeWarning:

divide by zero encountered in double_scalars

CV Minutes: 5.206169327100118

cv_df.head()

init = time()
cv_mlf_df = mlf.cross_validation(
    data=Y_df, 
    window_size=horizon, 
    n_windows=3, 
    step_size=horizon, 
    level=[90],
)
end = time()
print(f'CV Minutes: {(end - init) / 60}')

/home/ubuntu/miniconda/envs/statsforecast/lib/python3.10/site-packages/mlforecast/forecast.py:576: UserWarning:

Excuting `cross_validation` after `fit` can produce unexpected errors

/home/ubuntu/miniconda/envs/statsforecast/lib/python3.10/site-packages/mlforecast/forecast.py:468: UserWarning:

Please rerun the `fit` method passing a proper value to prediction intervals to compute them.

/home/ubuntu/miniconda/envs/statsforecast/lib/python3.10/site-packages/mlforecast/forecast.py:468: UserWarning:

Please rerun the `fit` method passing a proper value to prediction intervals to compute them.

/home/ubuntu/miniconda/envs/statsforecast/lib/python3.10/site-packages/mlforecast/forecast.py:468: UserWarning:

Please rerun the `fit` method passing a proper value to prediction intervals to compute them.

CV Minutes: 2.961174162228902

cv_mlf_df.head()

cv_nf_df = pd.read_parquet('https://m5-benchmarks.s3.amazonaws.com/data/cross-validation-nf.parquet')

cv_nf_df.head()

cv_df = cv_df.merge(cv_nf_df.drop(columns=['y']), how='left', on=['unique_id', 'ds', 'cutoff'])
cv_df = cv_df.merge(cv_mlf_df.drop(columns=['y']), how='left', on=['unique_id', 'ds', 'cutoff'])

cutoffs = cv_df['cutoff'].unique()

for cutoff in cutoffs:
    img = sf.plot(
        Y_df, 
        cv_df.query('cutoff == @cutoff').drop(columns=['y', 'cutoff']), 
        max_insample_length=28 * 5, 
        unique_ids=['FOODS_3_001_CA_1'],
    )
    img.show()

agg_cv_df = cv_df.loc[:,~cv_df.columns.str.contains('hi|lo')].groupby(['ds', 'cutoff']).sum(numeric_only=True).reset_index()
agg_cv_df.insert(0, 'unique_id', 'agg_demand')

agg_Y_df = Y_df.groupby(['ds']).sum(numeric_only=True).reset_index()
agg_Y_df.insert(0, 'unique_id', 'agg_demand')

for cutoff in cutoffs:
    img = sf.plot(
        agg_Y_df, 
        agg_cv_df.query('cutoff == @cutoff').drop(columns=['y', 'cutoff']),
        max_insample_length=28 * 5,
    )
    img.show()

from typing import List, Callable

from distributed import Client
from fugue import transform
from fugue_dask import DaskExecutionEngine
from datasetsforecast.losses import mse, mae, smape

def evaluate(df: pd.DataFrame, metrics: List[Callable]) -> pd.DataFrame:
    eval_ = {}
    models = df.loc[:, ~df.columns.str.contains('unique_id|y|ds|cutoff|lo|hi')].columns
    for model in models:
        eval_[model] = {}
        for metric in metrics:
            eval_[model][metric.__name__] = metric(df['y'], df[model])
    eval_df = pd.DataFrame(eval_).rename_axis('metric').reset_index()
    eval_df.insert(0, 'cutoff', df['cutoff'].iloc[0])
    eval_df.insert(0, 'unique_id', df['unique_id'].iloc[0])
    return eval_df

str_models = cv_df.loc[:, ~cv_df.columns.str.contains('unique_id|y|ds|cutoff|lo|hi')].columns
str_models = ','.join([f"{model}:float" for model in str_models])
cv_df['cutoff'] = cv_df['cutoff'].astype(str)
cv_df['unique_id'] = cv_df['unique_id'].astype(str)

client = Client() # without this, dask is not in distributed mode
# fugue.dask.dataframe.default.partitions determines the default partitions for a new DaskDataFrame
engine = DaskExecutionEngine({"fugue.dask.dataframe.default.partitions": 96})

evaluation_df = transform(
    cv_df.loc[:, ~cv_df.columns.str.contains('lo|hi')], 
    evaluate, 
    engine="dask",
    params={'metrics': [mse, mae, smape]}, 
    schema=f"unique_id:str,cutoff:str,metric:str, {str_models}", 
    as_local=True,
    partition={'by': ['unique_id', 'cutoff']}
)

/home/ubuntu/miniconda/envs/statsforecast/lib/python3.10/site-packages/distributed/client.py:3109: UserWarning:

Sending large graph of size 49.63 MiB.
This may cause some slowdown.
Consider scattering data ahead of time and using futures.

evaluation_df.head()

# Calculate the mean metric for each cross validation window
evaluation_df.groupby(['cutoff', 'metric']).mean(numeric_only=True)

!pip install seaborn

import matplotlib.pyplot as plt
import seaborn as sns

evaluation_df_melted = pd.melt(evaluation_df, id_vars=['unique_id', 'cutoff', 'metric'], var_name='model', value_name='error')

sns.violinplot(evaluation_df_melted.query('metric=="smape"'), x='error', y='model')

# Choose the best model for each time series, metric, and cross validation window
evaluation_df['best_model'] = evaluation_df.idxmin(axis=1, numeric_only=True)
# count how many times a model wins per metric and cross validation window
count_best_model = evaluation_df.groupby(['cutoff', 'metric', 'best_model']).size().rename('n').to_frame().reset_index()
# plot results
sns.barplot(count_best_model, x='n', y='best_model', hue='metric')

# For the mse, calculate how many times a model wins
eval_series_df = evaluation_df.query('metric == "mse"').groupby(['unique_id']).mean(numeric_only=True)
eval_series_df['best_model'] = eval_series_df.idxmin(axis=1)
counts_series = eval_series_df.value_counts('best_model')
plt.pie(counts_series, labels=counts_series.index, autopct='%.0f%%')
plt.show()

sf.plot(Y_df, cv_df.drop(columns=['cutoff', 'y']), 
        max_insample_length=28 * 6, 
        models=['AutoNHITS'],
        unique_ids=eval_series_df.query('best_model == "AutoNHITS"').index[:8])

# Merge the best model per time series dataframe
# and filter the forecasts based on that dataframe
# for each time series
fcst_df = pd.melt(fcst_df.set_index('unique_id'), id_vars=['ds'], var_name='model', value_name='forecast', ignore_index=False)
fcst_df = fcst_df.join(eval_series_df[['best_model']])
fcst_df[['model', 'pred-interval']] = fcst_df['model'].str.split('-', expand=True, n=1)
fcst_df = fcst_df.query('model == best_model')
fcst_df['name'] = [f'forecast-{x}' if x is not None else 'forecast' for x in fcst_df['pred-interval']]
fcst_df = pd.pivot_table(fcst_df, index=['unique_id', 'ds'], values=['forecast'], columns=['name']).droplevel(0, axis=1).reset_index()

sf.plot(Y_df, fcst_df, max_insample_length=28 * 3)