timeseries

timeseries crunches your data and selects the forecasting model that gives the best fit. You can let timeseries select the best model or specify the model you want. timeseries detects seasonality in your data. It considers periodic cycles when predicting what your data will look like in the future. You can specify the type of seasonality or let timeseries choose the best fit.

The amount of data, which is required to make a prediction depends on how your data is filtered and grouped. For example, for a non-seasonal monthly model, 2 data points are sufficient, whereas for a seasonal monthly model, at least 24 data points (two seasonal cycles) are required. If you don't have enough data to make a good prediction, timeseries returns nulls in the data. If no data is passed to timeseries, an empty dataset is returned.

1result = timeseries resultSet generate (measure1 as fmeasure1 [, measure2 as fmeasure2...]) with (parameters);

parameters can have the following values:

• length (required) Number of points to predict. For example, if length is 6 and the dateCols type string is Y-M, timeseries predicts data for 6 months.

  If you want to use dateCols but your data stream has missing dates, use fill before using timeseries.

  timeseries makes the most accurate prediction possible by choosing the best algorithm for your data. Predictive algorithms are more accurate for shorter time periods.

  Note

• dateCols (optional) Date fields to use for grouping the data, plus the date column type string. For example, dateCols=(CloseDate_Year, CloseDate_Month, "Y-M"). Date columns are projected automatically. Allowed values are:

  • YearField, MonthField, "Y-M"
  • YearField, QuarterField, "Y-Q"
  • YearField, "Y"
  • YearField, MonthField, DayField "Y-M-D"
  • YearField, WeekField "Y-W"

• ignoreLast (optional) If true, timeseries doesn't use the last time period in the calculations. The default is false.

  Set this parameter to true to improve the accuracy of the forecast if the last time period contains incomplete data. For example, if you’re partway through the quarter, timeseries forecasts more accurately if you set this parameter to true.

• order (optional) Specify the field to use for ordering the data. Mandatory if dateCols isn’t used. By default, this field is sorted in ascending order. Use desc to specify descending order, for example order=('Type' desc). You can also order by multiple fields, for example order=('Type' desc, 'Group' asc).

  For example, suppose that your data has no date columns, but it has a measure column called Week. Use order='Week'.

  Specify either dateCols or order.

  Note

• partition (optional) Specify the column used to partition the data. The column must be a dimension. The timeseries calculation is done separately for each partition to ensure that each partition uses the most accurate algorithm. For example, data in one partition might have a seasonal variation while data in another partition doesn't. The partition columns are projected automatically.

  For example, suppose that your sales data for raw materials contains the date sold, type of raw material, and the weight sold. To predict the future weight sold for each type of raw material, use partition='Type'.

• predictionInterval (optional) Specify the uncertainty, or confidence interval, to display at each point. Allowed values are 80 and 95. The upper and lower bounds of the confidence interval are projected in columns named column_name_low_95 and column_name_high_95.

• model (optional) Specify which prediction model to use. If unspecified, timeseries calculates the prediction for each model and selects the best model using Bayesian information criterion (BIC).

  Allowed values are:

  • None timeseries selects the best algorithm for the data
  • Additive uses Holt's Linear Trend or Holt-Winters method with additive components.
  • Multiplicative uses Holt's Linear Trend or Holt-Winters method with multiplicative components
• seasonality (optional) Use with dateCols to specify the seasonality for your prediction. Allowed values are:

  • 0 No seasonality
  • any integer between 2 and 24

  If unspecified, timeseries calculates the prediction for each type of seasonality and select the results with the smallest error.

  Example

  seasonality	dateCols	Type of Seasonality
  seasonality=4	dateCols="Y-Q"	Yearly seasonality, because there are four quarters in a year.
  seasonality=12	dateCols="Y-M"	Yearly seasonality, because there are 12 months in a year.
  seasonality=7	dateCols="Y-M-D"	Weekly seasonality, because there are seven days in a week.

Here's how you can make the most of using timeseries:

• Are you currently part way through the month, quarter, or year? Consider setting ignoreLast to true so that timeseries doesn't use the partial data in the current time period, leading to a more accurate prediction.
• Is timeseries not returning any data? If there aren't enough data points to make a good prediction, timeseries returns null. Try increasing the number of data points.
• Is timeseries returning an error? You could have gaps in your dates or times. Like all good forecasting algorithms, timeseries needs a continuous set of dates with no gaps, including in each partition. If you think your data has date gaps, try using fill first.