How to compute rolling weighted moving average as feature during training and prediction.

[anand086]

First of all, thank you for creating this wonderful library. I am working on a time-series problem to forecast for upcoming 21 days and was wondering how can I implement rolling weighted moving average. During feature engineering I have used the below code to calculate rolling weighted moving average, but I am not sure how can I provide during predict. My understanding is that when I go to make forecasts recursively, then I would need to compute the value of this features after making my predictions for each period.

window_size=7

# Define weighted mean function to pass to .apply()
def weighted_mean(x, weights):
    return (weights * x).sum() / weights.sum()

weights = np.arange(1, window_size + 1)  # Adjust window_size as needed

rolling_wma_7 = df_.groupby('unique_id')['y'].rolling(window=window_size).apply(weighted_mean, args=(weights,)).shift(1).reset_index(drop=True)
df_['rolling_wma_7'] = rolling_wma_7

[jmoralez]

Hey. This isn't really straightforward, but you can achieve it with the following:

import numpy as np
import pandas as pd
from mlforecast import MLForecast
from mlforecast.feature_engineering import transform_exog
from sklearn.linear_model import LinearRegression
from utilsforecast.data import generate_series

horizon = 10
window_size = 7
# create target
series = generate_series(1, min_length=100, max_length=100)
# create weights. note that you'll need the future values of these
weights = pd.DataFrame(
    {
        'unique_id': series.loc[0, 'unique_id'],
        'ds': pd.date_range(series['ds'].min(), series['ds'].max() + horizon * pd.offsets.Day()),
        'weight': np.random.rand(series.shape[0] + horizon)
    }
)
# assign the weight for each lag
lags = range(1, window_size + 1)
full_weights = transform_exog(weights, lags=lags).dropna().drop(columns='weight')
# merge to series. this drops the first observations because there aren't weights for that lag
train = series.merge(full_weights, on=['unique_id', 'ds'], how='inner')
fcst = MLForecast(
    models=[LinearRegression()],
    lags=lags,
    freq='D',
)
# create the lags
X, y = fcst.preprocess(train, static_features=[], return_X_y=True)

# define a function that will compute the weighted rolling mean by multiplying the lags and weights columns
def compute_features(df):
    df = df.copy(deep=False)    
    lag_cols = [f'lag{i+1}' for i in range(7)]
    weight_cols = [f'weight_lag{i+1}' for i in range(7)]
    lags = df[lag_cols].to_numpy()
    weights = df[weight_cols].to_numpy()
    df['rolling_wma_7'] = (lags * weights).sum(axis=1) / weights.sum(axis=1)
    df = df.drop(columns=lag_cols + weight_cols)
    return df

# apply this processing to the training set
prep_X = compute_features(X)
# train the models on that
fcst.fit_models(prep_X, y)
# provide the future weights and the processing function
fcst.predict(horizon, X_df=full_weights, before_predict_callback=compute_features)