kibana-123/src/core_plugins/timelion/server/series_functions/holt/index.js

Kibana is an open source (Apache Licensed), browser based analytics and search dashboard for Elasticsearch. Kibana is a snap to setup and start using. Kibana strives to be easy to get started with, while also being flexible and powerful, just like Elastic

var _ = require('lodash');
var Chainable = require('../../lib/classes/chainable');
var ses = require('./lib/ses');
var des = require('./lib/des');
var tes = require('./lib/tes');
var toMilliseconds = require('../../lib/to_milliseconds');

module.exports = new Chainable('holt', {
  args: [
    {
      name: 'inputSeries',
      types: ['seriesList']
    },
    {
      name: 'alpha',
      types: ['number'],
      help: `
        Smoothing weight from 0 to 1.
        Increasing alpha will make the new series more closely follow the original.
        Lowering it will make the series smoother`
    },
    {
      name: 'beta',
      types: ['number'],
      help: `
        Trending weight from 0 to 1.
        Increasing beta will make rising/falling lines continue to rise/fall longer.
        Lowering it will make the function learn the new trend faster`
    },
    {
      name: 'gamma',
      types: ['number'],
      help: `
        Seasonal weight from 0 to 1. Does your data look like a wave?
        Increasing this will give recent seasons more importance, thus changing the wave form faster.
        Lowering it will reduce the importance of new seasons, making history more important.
        `
    },
    {
      name: 'season',
      types: ['string'],
      help: 'How long is the season, eg, 1w if you pattern repeats weekly. (Only useful with gamma)'
    },
    {
      name: 'sample',
      types: ['number', 'null'],
      help: `
      The number of seasons to sample before starting to "predict" in a seasonal series.
      (Only useful with gamma, Default: all)`
    }
  ],
  help: `
    Sample the beginning of a series and use it to forecast what should happen
    via several optional parameters. In general, like everything, this is crappy at predicting the
    future. You're much better off using it to predict what should be happening right now, for the
    purpose of anomaly detection. Note that nulls will be filled with forecasted values. Deal with it.`,
  fn: function expsmoothFn(args, tlConfig) {

    let newSeries = _.cloneDeep(args.byName.inputSeries);

    const alpha = args.byName.alpha;
    const beta = args.byName.beta;
    const gamma = args.byName.gamma;

    _.each(newSeries.list, function (series) {
      const sample = args.byName.sample || series.data.length; // If we use length it should simply never predict


      // Single exponential smoothing
      // This is basically a weighted moving average in which the older
      // points exponentially degrade relative to the alpha, eg:
      // 0.8^1, 0.8^2, 0.8^3, etc

      var times = _.map(series.data, 0);
      var points = _.map(series.data, 1);

      if (alpha != null && beta == null && gamma == null) {
        points = ses(points, alpha);
      }

      if (alpha != null && beta != null && gamma == null) {
        points = des(points, alpha, beta);
      }

      if (alpha != null && beta != null && gamma != null) {
        if (!sample || !args.byName.season || sample < 2) {
          throw new Error('Must specificy a season length and a sample size >= 2');
        }
        const season = Math.round(toMilliseconds(args.byName.season) / toMilliseconds(tlConfig.time.interval));
        points = tes(points, alpha, beta, gamma, season, sample);
      }

      _.assign(series.data, _.zip(times, points));
    });

    return newSeries;
  }
});