fcompdata

Forecasting Competitions Datasets - a Python library for loading M and tourism competitions time series datasets (M1, M3, M4, Tourism) with an interface similar to R's Mcomp and Tcomp packages.

Installation

pip install fcompdata

or from github:

pip install git+https://github.com/config-i1/fcompdata

Usage

from fcompdata import M1, M3, Tourism

# Access series by 1-based index (R-style)
series = M3[1]
print(series['x'])    # Training data (numpy array)
print(series['xx'])   # Test data (numpy array)
print(series['y'])    # Full series: concat(x, xx), length n + h
print(series['h'])    # Forecast horizon
print(series['n'])    # Training data length
print(series['type']) # Series type (yearly, quarterly, monthly, other)

# Attribute access also works
print(series.sn)          # Series name
print(series.description) # Series description

# Filter by frequency type
yearly = M3.subset('yearly')
monthly = M1.subset('monthly')

# Iterate over all series
for series in M3:
    print(series.sn, len(series.x))

# Get series count
print(len(M3))  # 3003

M4 Dataset

The M4 competition dataset contains 100,000 time series and is too large to bundle with the package. It must be downloaded separately before use. The data is sourced from the Monash Time Series Forecasting Repository hosted on Zenodo.

Downloading M4 Data

from fcompdata.download import download_m4

# Download all M4 frequencies (~50MB total, saved to ~/.fcompdata/m4/)
download_m4()

# Or download specific frequencies
download_m4('yearly')     # 23,000 series
download_m4('quarterly')  # 24,000 series
download_m4('monthly')    # 48,000 series
download_m4('weekly')     # 359 series
download_m4('daily')      # 4,227 series
download_m4('hourly')     # 414 series

The data is downloaded once and cached locally in ~/.fcompdata/m4/. Subsequent calls will use the cached files.

Using M4 Data

from fcompdata import M4, load_m4

# Load all M4 series (requires all frequencies to be downloaded)
series = M4[1]

# Load a specific frequency
yearly = load_m4('yearly')
monthly = load_m4('monthly')

# Same interface as other datasets
print(series.x)       # Training data
print(series.xx)      # Test data
print(series.h)       # Forecast horizon
print(series.type)    # 'yearly', 'quarterly', etc.

# Filter and iterate
for s in yearly:
    print(s.sn, len(s.x))

M4 Download Sources

The M4 data files are downloaded from the Monash Time Series Forecasting Repository on Zenodo:

Cache Management

from fcompdata.download import clear_cache, get_m4_path

# Check if a frequency is downloaded
path = get_m4_path('yearly')  # Returns Path or None

# Clear all downloaded data
clear_cache()

# Clear only M4 data
clear_cache('m4')

Individual Time Series

In addition to the competition datasets, fcompdata bundles several classic individual time series ported from base R and the forecast package. These are tiny, load instantly, and behave like a single MCompSeries (x, xx, h, period, type, description). Two of them carry exogenous regressors on xreg / xregx / xregxx.

from fcompdata import AirPassengers, BJsales, Seatbelts, taylor

# Same MCompSeries interface as the competition series
print(AirPassengers.x)         # 132 training observations
print(AirPassengers.xx)        # 12 holdout observations
print(AirPassengers.period)    # 12 (monthly)

# Series with exogenous regressors are stored as numpy structured arrays
# (recarray), so the column names of explanatory variables are preserved:
print(BJsales.xreg.dtype.names)            # ('BJsales.lead',)
print(BJsales.xreg['BJsales.lead'][:5])    # first five values

print(Seatbelts.xreg.dtype.names)          # ('kms', 'PetrolPrice', 'law')
print(Seatbelts.xreg.kms[:5])              # 1-D array, recarray attribute access
print(Seatbelts.xregxx['law'])             # last 12 values of the law indicator

# xreg is the row-wise concatenation of xregx (training) and xregxx (holdout).
# To get a plain 2-D float matrix for linear algebra:
import numpy as np
mat = np.column_stack([Seatbelts.xreg[n] for n in Seatbelts.xreg.dtype.names])

Note: BJsales and BJsales.lead have frequency=1 in R. fcompdata stores them with period=12 and type='monthly' to match the requested holdout of twelve observations; the original R metadata is documented in the series description.

Datasets

Bundled Datasets

These datasets are included with the package and available immediately:

Downloadable Datasets

These datasets require downloading before use:

Series Attributes

Each MCompSeries object has the following attributes:

Data Sources

The time series data in this package was imported from the following sources:

• Mcomp (M1 and M3 data): Hyndman, R.J. (2024). Mcomp: Data from the M-Competitions. R package. CRAN, GitHub
• Tcomp (Tourism data): Hyndman, R.J. (2016). Tcomp: Data from the 2010 Tourism Forecasting Competition. R package. CRAN, GitHub
• Monash Time Series Forecasting Repository (M4 data): forecastingdata.org, hosted on Zenodo
• R datasets package (AirPassengers, BJsales, BJsales.lead, Seatbelts): bundled with base R. CRAN
• R forecast package (taylor): Hyndman, R.J. (2024). forecast: Forecasting functions for time series and linear models. R package. CRAN, GitHub
• CMAF Demand Forecasting course (PromoData): Svetunkov, I. (2024). Demand Forecasting course materials (Session 6.2 — ETS with regressors). Centre for Marketing Analytics and Forecasting (CMAF), Lancaster University Management School.

References

The datasets were used in the following forecasting competitions:

M1 Competition:

Makridakis, S., Andersen, A., Carbone, R., Fildes, R., Hibon, M., Lewandowski, R., Newton, J., Parzen, E., & Winkler, R. (1982). The accuracy of extrapolation (time series) methods: Results of a forecasting competition. Journal of Forecasting, 1(2), 111–153. doi:10.1002/for.3980010202

M3 Competition:

Makridakis, S., & Hibon, M. (2000). The M3-Competition: Results, conclusions and implications. International Journal of Forecasting, 16(4), 451–476. doi:10.1016/S0169-2070(00)00057-1

M4 Competition:

Makridakis, S., Spiliotis, E., & Assimakopoulos, V. (2020). The M4 Competition: 100,000 time series and 61 forecasting methods. International Journal of Forecasting, 36(1), 54–74. doi:10.1016/j.ijforecast.2019.04.014

Tourism Forecasting Competition:

Athanasopoulos, G., Hyndman, R.J., Song, H., & Wu, D.C. (2011). The tourism forecasting competition. International Journal of Forecasting, 27(3), 822–844. doi:10.1016/j.ijforecast.2010.11.005

Monash Time Series Forecasting Archive:

Godahewa, R., Bergmeir, C., Webb, G.I., Hyndman, R.J., & Montero-Manso, P. (2021). Monash Time Series Forecasting Archive. Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks (NeurIPS Datasets and Benchmarks 2021). arXiv:2105.06643

The individual time series come from the following original sources:

AirPassengers:

Box, G. E. P., Jenkins, G. M., Reinsel, G. C., & Ljung, G. M. (2015). Time Series Analysis, Forecasting and Control (5th ed.). Wiley. Series G.

BJsales / BJsales.lead:

Box, G. E. P., & Jenkins, G. M. (1976). Time Series Analysis, Forecasting and Control. Holden-Day. Series M.

Seatbelts:

Harvey, A. C., & Durbin, J. (1986). The effects of seat belt legislation on British road casualties: A case study in structural time series modelling. Journal of the Royal Statistical Society A, 149, 187–227. doi:10.2307/2981553

taylor:

Taylor, J. W. (2003). Short-term electricity demand forecasting using double seasonal exponential smoothing. Journal of the Operational Research Society, 54, 799–805. doi:10.1057/palgrave.jors.2601589

PromoData:

Svetunkov, I. (2024). Demand Forecasting course materials (Session 6.2 — ETS with regressors). Centre for Marketing Analytics and Forecasting (CMAF), Lancaster University Management School.

License

LGPL-3.0-or-later

About

fcompdata is developed and maintained by OpenForecast, a demand forecasting and inventory management consultancy. The package implements the methods we use in our consulting and teach in our training courses.