fcst 0.8.3

Forecasting utilities, time-series preprocessing, rolling back-test and ensemble forecasting from model selection

fcst

Package repo on PyPI: fcst - PyPI

Installation

uv add fcst

Features

This package provides you with these sub-modules

1. automation

   This automatically runs back-test, select the best models, and forecast for you. You can customise whether or not to run in parallel, how many top models to select, etc.

2. forecasting

   This provides you with the basic functionality of fit() and predict(), given that you pass in the model.

3. evaluation

   This provides you with back-test and model selection functionalities.

4. preprocessing

   This allows you to prepare your dataframes, preprocess the time-series data, fill in the missing dates automatically.

5. horizon

   This is an API for dealing with future horizon from sktime. But in some modules, it will also do this automatically.

6. models

   Gives you the base models for you to work with. Provides you with the basic models, default (fallback) and zero predictor.

7. metrics

   Our own implementation of forecasting performance metrics.

8. common

   Other common functionalities, e.g., types.

Usage

Examples

In case you want to automate the whole process...

from fcst.automation import run_forecasting_automation
import pandas as pd


df_input = pd.read_csv("path-to-your/file.csv")

data_period_date = pd.Period("2025-02", freq="M")

# Can support multivariate forecasting
# Just pass in the config
multivar_config = MultiVarConfig(df_X_raw=df_X, feature_cols=["feature_1", "feature_2", "feature_3"])

results = run_forecasting_automation(
    df_raw=df_y,
    value_col="y",
    date_col="date",
    data_period_date=data_period_date,
    forecasting_periods=5,
    id_cols=["id"],
    multivar_config=multivar_config,
)

# Do something with the results
def format_and_upload_results(df_results):
    ...

There are main configs you can customise...

from fcst.common.configs import ForecastingConfig, BacktestingConfig

forecasting_config = ForecastingConfig(top_n=2)
backtesting_config = BacktestingConfig(backtesting_periods=3, eval_periods=4)


df_input = pd.read_csv("path-to-your/file.csv")

data_period_date = pd.Period("2025-02", freq="M")

results = run_forecasting_automation(
    df_raw=df_y,
    value_col="y",
    date_col="date",
    data_period_date=data_period_date,
    forecasting_periods=5,
    id_cols=["id"],
    forecasting_config=forecasting_config,
    backtesting_config=backtesting_config
)

# Do something with the results
def format_and_upload_results(df_results):
    ...

Other utilities for time-series

from fcst.preprocessing import prepare_timeseries


# Group time-series based on customer-product, then yields a generater of (id_, pd.Series)
timeseries = prepare_timeseries(
    df_raw=df_input,
    date_col="date",
    value_col="net_amount",
    data_period_date=data_period_date,
    id_cols=["customer_code", "product_code"],
    min_cap=0,
    freq="M",
    agg_method="sum",
    fillna=0,
)


# Returns the whole DataFrame as a time-series
timeseries = prepare_timeseries(
    df_raw=df_input,
    date_col="date",
    value_col="net_amount",
    data_period_date=data_period_date,
    id_cols=None,
    min_cap=0,
    freq="M",
    agg_method="sum",
    fillna=0,
)

More detailed usage

Automation

This automation sub-module runs cleaning, fill-in the missing dates, evaluation, model selection, forecasting and ensemble, everything automatically.

Preprocessing

This provides time-series preparation functions. The most complete function and does most of the heavy lifting is prepare_timeseries() function. You can import it from from fcst.preprocessing import prepare_timeseries. You can specify ID columns, date column, value column to forecast, the frequency, and much more. This function returns a dictionary where the keys are time-series IDs and the values are the corresponding time-series.

Models

By default the run_forecasting_automation() uses base_models from models. But you can define your own model(s) with fit() and predict() methods. You can get the base_models, fast_models, all_models, or multivar_models and put your own model(s) to the dictionary.

Metrics

By default, we use smape() for measuring accuracy (error) of forecasting models. We define our own as it handles when the forecast or actual values are 0. There are other metrics as well such as: - mape() - mae() - mse() - rmse()

Horizon

If you want to utilise the forecast horizons or get some future dates, this sub-module provides the basic functionalities. And this is the base of forecasting and evaluation functions.

Forecasting

In forecasting sub-module, you can use forecast or ensemble to forecast using one or more models.

Evaluation

The evaluation sub-module provides the back-testing and model selection functions. You can pass in a model dictionary to evaluate which models are suitable for each time-series.