JoaquinAmatRodrigo
commented
1 year ago Hi @JorgeGomes72 , thanks for using skforecast. Your use case is what we call an intermittent demand with a regular pattern. This is a field we are actively exploring, and we will soon publish a user's guide that will discuss it.
Why does your model predict non-zero values? I see that the predicted values for Sundays are much lower than predictions for other days, yet the model is not able to learn that it should be exactly zero. Since you mention that there are no Sundays in the historical data with sales, the problem is not in the training data. I think the model has learned that sales on one day are highly correlated with sales on previous days. This is true for all days of the week except Sundays. So the model is torn between learning a general pattern and a local one.
How to solve it.
There are some options. If what is really important for your business case is the predictions for all days except Sunday, try to find the model that best predicts the days Monday through Saturday (you can use a custom metric to ignore specific values). Once you have the predictions, replace all Sunday values with zero.
If you really need the output of the model to be always positive, you can apply a log transformation to the data before fitting the model.
I hope this helps!
JorgeGomes72
Hello, I'm building a time series model, for forecasting Sales, based on the history since 2019. My model uses SKForecast with XGBRegressor.
I want to forecast 75 days. My target is SALES.
I use external features to help model, transformed in 0 and 1.
I would like to understand why my final forecast have non-zero values on target variable SALES on Sundays, even when i use a external feature OPEN=0 and even when i have SALES=0 on every Sundays in the history.
My dataset have this struture: DATA| SALES| YEAR | WEEK | WEEKDAY | OPEN
ex:
The variable "WEEKDAY"= 7 means Sunday. On dataset i have every Sundays with SALES=0, with OPEN=0.
The external feature "OPEN"=0 means that store is closed, OPEN=1 means store open.
This is my final dataset (vendas_df2) before execute model :
The exog_variables uses all the external features, except target variable (SALES)
This is the train, validation and test :
This is the parameters for model:
Create forecaster
======================================
forecaster = ForecasterAutoreg( regressor = XGBRegressor(random_state=123), lags = 7 #24 )
Grid search of hyperparameters and lags
========================================
Regressor hyperparameters
param_grid = { 'n_estimators': [100, 500], 'max_depth': [3, 5, 10], 'learning_rate': [0.01, 0.1] }
Lags used as predictors
lags_grid = [7, 30, 48, 72, [1, 2, 3, 7, 23, 24, 25, 71, 72, 73]]
results_grid = grid_search_forecaster( forecaster = forecaster, y = vendas_df2.loc[:end_validation, 'SALES'], exog = vendas_df2.loc[:end_validation, exog_variables],
param_grid = param_grid, lags_grid = lags_grid, steps = 75, refit = False, metric = 'mean_squared_error', initial_train_size = int(len(data_train)), fixed_train_size = False, return_best = True, verbose = False )
Backtesting test data
=========================================
metric, predictions = backtesting_forecaster( forecaster = forecaster, y = vendas_df2['SALES'], exog = vendas_df2[exog_variables], initial_train_size = len(vendas_df2.loc[:end_validation]), fixed_train_size = False, steps = 75, refit = False, metric = 'mean_squared_error', verbose = False )
print(f"Backtest error: {metric}")
This is the final result with forecast for May:
We can see that Sundays have SALES <>0.
My dataset has OPEN=0 for Sundays, so why i can't forecast zero values for Sundays (prev=0) ?
Can you help please? Thank you!
Jorge Gomes
Originally posted by @JorgeGomes72 in https://github.com/JoaquinAmatRodrigo/skforecast/discussions/388