com.yahoo.egads.models.tsmm

Class OlympicModel2

java.lang.Object

com.yahoo.egads.models.tsmm.TimeSeriesAbstractModel

com.yahoo.egads.models.tsmm.OlympicModel2

All Implemented Interfaces:

JsonAble, Model, TimeSeriesModel, java.io.Serializable

public class OlympicModel2
extends TimeSeriesAbstractModel

TODO - rename this! This model generates forecasts using an average of period-over-period baseline data. For example, if the window distance is 1 week and the user asks for 4 past windows, a prediction at time T will be generated from the average (or WINDOW_AGGREGATION) of T - 1 week, T - 2 weeks, T - 3 weeks and T - 4 weeks.

Properties

• INTERVAL - REQUIRED: The timestamp interval of the resulting model, in combination with the PERIOD_SIZE_UNITS. E.g. for a 5 minute window period, set this value to "5".
• INTERVAL_UNITS - REQUIRED: A Java ChronoUnit enumerator representing the units of the PERIOD_SIZE. E.g. "minutes" or "seconds".
• WINDOW_SIZE - REQUIRED: How many values should be generated for the model, e.g. if you want to generate a model for a day, set this to 1 and set WINDOW_SIZE_UNITS to "days"
• WINDOW_SIZE_UNITS - REQUIRED: A Java ChronoUnit enumerator representing the units of the WINDOW_SIZE. E.g. "minutes" or "seconds".
• WINDOW_DISTANCE - The timestamp interval between windows in the baseline. E.g. for a week-over-week model, set this value to "1".
• WINDOW_DISTANCE_UNITS - REQUIRED: A Java ChronoUnit enumerator representing the units of the WINDOW_DISTANCE. E.g. "minutes" or "seconds".
• MODEL_START - REQUIRED: The start time of the model in Unix epoch seconds. This will be the first timestamp of the model and determines the offsets for parsing the baseline. E.g. 1474756200.
• HISTORICAL_WINDOWS - The number of windows in the past to generate the model from. E.g. for 4 weeks of week-over-week, set this to 4. Defaults to 1 window.
• WINDOW_AGGREGATOR - An aggregation function to use when merging data points across historical windows. Can be one of SUM, MAX, MIN, COUNT, AVG, WAVG or MEDIAN. Defaults to AVG.
• TIMEZONE - An optional time zone to align data to. Must be parseable by Java's ZoneId, e.g. "Australia/Lord_Howe". Defaults to "UTC".
• ENABLE_WEIGHTING - Whether or not to enable higher weighting for more recent values in the baseline. Defaults to false.
• FUTURE_WINDOWS - How many periods to forecast in the future for generating the model. Defaults to 1.
• NUM_TO_DROP_HIGHEST - How many of the highest values to drop amongst the windows at each timestamp. Defaults to 0.
• NUM_TO_DROP_LOWEST - How many of the lowest values to drop amongst the windows at each timestamp. Defaults to 0.

Optionally weighting can be applied wherein more recent values (e.g. T - 1 week) have more weight than older values (e.g. T - 4 weeks).

Also, for each data point, the n highest and n lowest values can be dropped. Using the example above, if we have T - 1w = 4, T - 2w = 2, T - 3w = 5, T - 4w = 1 and we drop 1 high and 1 low, then the resulting would be the average of T - 1w and T - 2w for a value of 2.

If a time zone is given, the baseline and forecast timestamps will be adjusted for DST if necessary.

See Also:

Serialized Form

Constructor Summary

Constructors

Constructor and Description
OlympicModel2(java.util.Properties config) Default Ctor

Method Summary

Modifier and Type	Method and Description
void	predict(TimeSeries.DataSequence sequence)
void	reset()
void	train(TimeSeries.DataSequence data)
void	update(TimeSeries.DataSequence data)

Methods inherited from class com.yahoo.egads.models.tsmm.TimeSeriesAbstractModel

betterThan, fromJson, getBias, getMAD, getMAPE, getModelName, getModelType, getMSE, getSAE, toJson

Methods inherited from class java.lang.Object

equals, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Detail

OlympicModel2

public OlympicModel2(java.util.Properties config)

Default Ctor

Parameters:

config - A non-null and non-empty properties map.

Throws:

java.lang.IllegalArgumentException - if a required property is missing.

java.lang.NumberFormatException - if a numeric property could not be parsed.

Method Detail

train

public void train(TimeSeries.DataSequence data)
           throws java.lang.Exception

Throws:

java.lang.Exception

update

public void update(TimeSeries.DataSequence data)
            throws java.lang.Exception

Throws:

java.lang.Exception

predict

public void predict(TimeSeries.DataSequence sequence)
             throws java.lang.Exception

Throws:

java.lang.Exception

reset

public void reset()