++ed by:
CHENRYN TSIBLEY

2 PAUSE users

Philipp K. Janert
and 1 contributors

NAME

Math::HoltWinters - Time series smoothing and forecasting using exponential smoothing

SYNOPSIS

  use Math::HoltWinters;

  $s = Math::HoltWinters::single( $alpha );
  $s = Math::HoltWinters::single( $alpha, $x0 );

  $s = Math::HoltWinters::double( $alpha, $beta );
  $s = Math::HoltWinters::double( $alpha, $beta, $x0 );
  $s = Math::HoltWinters::double( $alpha, $beta, $x0, $t0 );

  $s = Math::HoltWinters::triple_add( $alpha, $beta, $gamma, $n );
  $s = Math::HoltWinters::triple_add( $alpha, $beta, $gamma, $x0, $n );
  $s = Math::HoltWinters::triple_add( $alpha, $beta, $gamma, $x0, $t0, $n );

  $s = Math::HoltWinters::triple_add( $alpha, $beta, $gamma, \@p );
  $s = Math::HoltWinters::triple_add( $alpha, $beta, $gamma, $x0, \@p );
  $s = Math::HoltWinters::triple_add( $alpha, $beta, $gamma, $x0, $t0, \@p );

  $s = Math::HoltWinters::triple_mul( $alpha, $beta, $gamma, $n );
  $s = Math::HoltWinters::triple_mul( $alpha, $beta, $gamma, $x0, $n );
  $s = Math::HoltWinters::triple_mul( $alpha, $beta, $gamma, $x0, $t0, $n );

  $s = Math::HoltWinters::triple_mul( $alpha, $beta, $gamma, \@p );
  $s = Math::HoltWinters::triple_mul( $alpha, $beta, $gamma, $x0, \@p );
  $s = Math::HoltWinters::triple_mul( $alpha, $beta, $gamma, $x0, $t0, \@p );


  # Smoothing data:
  for( @data ) {
    push @smoothed, $s->( $_ );
  }

  # Forecasting five steps:
  for( 1..5 ) {
    push @forecast, $s->();
  }

  # Alternative syntax (including both smoothing and forecasting):
  @result = map { $s->( $_ ) } ( @data, (undef) x 5 );

DESCRIPTION

This module provides functions to perform exponential smoothing and forecasting for time series data (Holt-Winters method). The module supports single (for stationary time series without trend), double (for time series with trend) and triple (for time series with trend and seasonality) exponential smoothing. Separate methods exist to handle additive and multiplicative seasonality.

This module provides four functions (one for each form of exponential smoothing), which instantiate a function reference. When applied to the raw time series data, this function reference will return the smoothed value of the time series; when applied to an undefined argument, the function reference will return a forecast.

Caveat: The function references maintain state between invocations! (This is how they do their job). It is therefore necessary to let them operate on the input data only once and in the proper order. The function references can not be reused - if you want to redo a calculation (for instance with different values for the smoothing parameters), you must instantiate new function references.

DETAILS

All functions in this module return a function reference, which can be used for smoothing and forecasting. All functions take between one and three smoothing parameters ($alpha, $beta, $gamma), which control the amount of smoothing applied. These parameters are mandatory, and their values should (but are not required to) fall between 0 and 1. All functions also take optional values which are used to start up the recursion; if they are not supplied, appropriate values are inferred from the data. (These hints are typically only necessary for very short data sets.)

  $s = Math::HoltWinters::single( $alpha );
  $s = Math::HoltWinters::single( $alpha, $x0 );

Instantiates a function reference that performs single exponential smoothing, with smoothing parameter $alpha. The value to be used for the initial smoothed point can be supplied as an optional parameter, if it not provided, the initial data point is used as the initial smoothed value.

  $s = Math::HoltWinters::double( $alpha, $beta );
  $s = Math::HoltWinters::double( $alpha, $beta, $x0 );
  $s = Math::HoltWinters::double( $alpha, $beta, $x0, $t0 );

Instantiates a function reference that performs double exponential smoothing, with smoothing parameter $alpha and $beta. Values for the initial smoothed point and the initial smoothed trend can be supplied as optional parameters. If they are not provided, they are calculated from the first points of the data.

  $s = Math::HoltWinters::triple_add( $alpha, $beta, $gamma, $n );
  $s = Math::HoltWinters::triple_add( $alpha, $beta, $gamma, $x0, $n );
  $s = Math::HoltWinters::triple_add( $alpha, $beta, $gamma, $x0, $t0, $n );

  $s = Math::HoltWinters::triple_add( $alpha, $beta, $gamma, \@p );
  $s = Math::HoltWinters::triple_add( $alpha, $beta, $gamma, $x0, \@p );
  $s = Math::HoltWinters::triple_add( $alpha, $beta, $gamma, $x0, $t0, \@p );

Instantiates a function reference that performs triple exponential smoothing, assuming additive seasonality, with smoothing parameter $alpha, $beta, and $gamma. Values for the initial smoothed point and the initial smoothed trend can be supplied as optional parameters. If they are not provided, they are calculated from the first points of the data.

The number of points per season must be provided for triple exponential smoothing; this information is always supplied through the last argument to the constructor. There are two ways to supply this seasonality information: if the last argument is a scalar, it is interpreted as the number of points per season and the remaining information is inferred from the data. Alternatively, the last argument can be reference to an array of the appropriate length (that is, as many elements as there are points in a season) holding initial values for the magnitude of the seasonality effect.

  $s = Math::HoltWinters::triple_mul( $alpha, $beta, $gamma, $n );
  $s = Math::HoltWinters::triple_mul( $alpha, $beta, $gamma, $x0, $n );
  $s = Math::HoltWinters::triple_mul( $alpha, $beta, $gamma, $x0, $t0, $n );

  $s = Math::HoltWinters::triple_mul( $alpha, $beta, $gamma, \@p );
  $s = Math::HoltWinters::triple_mul( $alpha, $beta, $gamma, $x0, \@p );
  $s = Math::HoltWinters::triple_mul( $alpha, $beta, $gamma, $x0, $t0, \@p );

Instantiates a function reference that performs triple exponential smoothing, assuming multiplicative seasonality, with smoothing parameter $alpha, $beta, and $gamma. Values for the initial smoothed point and the initial smoothed trend can be supplied as optional parameters. If they are not provided, they are calculated from the first points of the data.

The number of points per season must be provided for triple exponential smoothing; this information is always supplied through the last argument to the constructor. There are two ways to supply this seasonality information: if the last argument is a scalar, it is interpreted as the number of points per season and the remaining information is inferred from the data. Alternatively, the last argument can be reference to an array of the appropriate length (that is, as many elements as there are points in a season) holding initial values for the magnitude of the seasonality effect.

USAGE

The function reference returned from any of the four functions can be applied to any numeric value, returning a numeric (smoothed) value. If the function reference is invoked with an undefined argument (or none), it returns the best forecast, based on its most recent state. This feature can be used to extend a smoothed time series past the last data point.

  # Smoothing:
  for( @data ) {
    push @smoothed, $s->( $_ );
  }

  # Forecasting five steps:
  for( 1..5 ) {
    push @forecast, $s->();
  }

Because the function reference maintains state between invocations, it must be invoked exactly once for each data point, and the data points must be supplied in proper time order (from earliest to latest). Similarly, to create a forecast, the function reference has to be invoked (with undefined argument) immediately after it has been applied to the available data.

It is not possible to make the function reference go backward in time, or to reuse it for a second smoothing run. Instead, create a new function reference from scratch.

However, it is possible to have several function references in existence and operating concurrently - no state is shared across instances.

EXPORT

This module does not export any functions.

LIMITATIONS AND RATIONALE

The functions provided by this module only implement the Holt-Winters methods for smoothing and forecasting. They do neither provide functionality to evaluate the error between the smoothed and the raw data, nor to determine the "optimal" values of the smoothing parameters (alpha, beta, gamma).

The fitting parameters (alpha, beta, gamma) need to specified at initialization time and cannot be changed later. This is intentional, for two reasons: first of all, it corresponds to the typical use case of Holt-Winters methods (it is rare to change these parameters in the middle of a data set). More importantly, it serves as a reminder that the (stateful) function references cannot be reused: if you want to change the parameters, you have to obtain a new function reference.

MATHEMATICAL REFERENCE

The exponential smoothing calculations can be defined in several different ways. This module uses the following conventions, where d[i] is the raw data at time step i, and y[i+k] is the returned value (smoothed or forecasted) at time step i+k, and n is the number of points per season:

  # Single
  x[i] = alpha*d[i] + (1-alpha)*x[i-1]
  y[i+k] = x[i]                k=0, 1, 2, ...

  # Double
  x[i] = alpha*d[i] + (1-alpha)*(x[i-1] + t[i-1])
  t[i] = beta*(x[i] - x[i-1]) + (1-beta)*t[i-1]
  y[i+k] = x[i] + k*t[i]       k=0, 1, 2, ...

  # Triple, additive
  x[i] = alpha*(d[i] - p[i-n]) + (1-alpha)*(x[i-1] + t[i-1])
  t[i] = beta*(x[i] - x[i-1]) + (1-beta)*t[i-1]
  p[i] = gamma*(d[i] - s[i]) + (1-gamma)*p[i-n]
  y[i+k] = x[i] + k*t[i] + p[i-n+k]

  # Triple, multiplicative
  x[i] = alpha*d[i]/p[i-n] + (1-alpha)*(x[i-1] + t[i-1])
  t[i] = beta*(x[i] - x[i-1]) + (1-beta)*t[i-1]
  p[i] = gamma*d[i]/s[i] + (1-gamma)*p[i-n]
  y[i+k] = (x[i] + k*t[i])*p[i-n+k]

SEE ALSO

  • Data Analysis with Open Source Tools by Philipp K. Janert; O'Reilly, 2010 For a general introduction to data analysis. Time series analysis, including Holt-Winters methods, are treated in chapter 4.

  • The Analysis of Time Series: An Introduction by Chris Chatfield; Chapman & Hall, 6th ed, 2003 A more in-depth, yet practical and accessible introduction to time series analysis.

  • NIST/SEMATECH e-Handbook of Statistical Methods (http://www.itl.nist.gov/div898/handbook/index.htm) An online reference to statistical methods; section 6.4.3 introduces Holt-Winters methods.

AUTHOR

Philipp K. Janert, <janert at ieee dot org>, http://www.beyondcode.org

COPYRIGHT AND LICENSE

Copyright (C) 2011 by Philipp K. Janert

This library is free software; you can redistribute it and/or modify it under the same terms as Perl itself, either Perl version 5.10.0 or, at your option, any later version of Perl 5 you may have available.