One-pass approach to automated timeseries forecasting
Herein are timeseries preprocessing, model selection, and hyperparameter tuning techniques for forecasting development based on temporal statistics of a timeseries and a single feed-forward pass through a machine learning (ML) pipeline. In an embodiment, a computer hosts and operates the ML pipeline that automatically measures temporal statistic(s) of a timeseries. ML algorithm selection, cross validation, and hyperparameters tuning is based on the temporal statistics of the timeseries. The result from the ML pipeline is a rigorously trained and production ready ML model that is validated to have increased accuracy for multiple prediction horizons. Based on the temporal statistics, efficiency is achieved by asymmetry of investment of computer resources in the tuning and training of the most promising ML algorithm(s). Compared to other approaches, this ML pipeline produces a more accurate ML model for a given amount of computer resources and consumes fewer computer resources to achieve a given accuracy.
1 . A method comprising:
automatically measuring, for a timeseries, a seasonality and a stationarity differencing count;
assigning the stationarity differencing count to a respective differencing hyperparameter of each machine learning (ML) algorithm from a plurality of ML algorithms;
selecting, based on said assigning, a most accurate ML algorithm from the plurality of ML algorithms;
identifying a values subrange of a seasonality hyperparameter of the most accurate ML algorithm;
automatically selecting, based on said stationarity differencing count, a minimum length of a training subsequence of the timeseries;
tuning, based on the values subrange of the seasonality hyperparameter of the most accurate ML algorithm, hyperparameters of said most accurate ML algorithm; and
training, with a plurality of subsequences of the timeseries that have at least the minimum length, an ML model that is based on; said tuning and the most accurate ML algorithm;
wherein the method is performed by one or more computers.
2 . The method of claim 1 wherein:
the plurality of ML algorithms consists of said most accurate ML algorithm and a plurality of other ML algorithms;
the method does not comprise tuning hyperparameters of at least one ML algorithm of the plurality of other ML algorithms.
3 . The method of claim 1 further comprising validating, based on said stationarity differencing count, said most accurate ML algorithm during at least one selected from the group consisting of:
said selecting said most accurate ML algorithm and
said tuning said hyperparameters of said most accurate ML algorithm.
4 . The method of claim 3 wherein:
said validating said most accurate ML algorithm comprises automatically calculating a forecast horizon duration.
5 . The method of claim 3 wherein:
said validating said most accurate ML algorithm comprises measuring, for each seasonality of a plurality of seasonalities, a respective fitness of said most accurate ML algorithm;
said fitnesses of said most accurate ML algorithm consists of:
a) a highest fitness for a particular seasonality of said plurality of seasonalities and
b) fitnesses for other seasonalities;
said tuning said hyperparameters of said most accurate ML algorithm is based on said particular seasonality and not said other seasonalities.
6 . The method of claim 1 further comprising validating the ML model based on said stationarity and at least one selected from the group consisting of seasonality and frequency.
7 . The method of claim 1 further comprising
measuring seasonality by at least one selected for the group consisting of:
fitting a second-degree polynomial regression,
fitting a polynomial regression with an ordinary least squares estimator,
applying an autocorrelation function (ACF) to a differenced timeseries that is based on said timeseries,
calculating a coefficient at specific lagged values of a differenced timeseries that is based on said timeseries,
calculating a coefficient that represent a strength of a linear relationship between respective values in two seasons of a same seasonality,
sorting respective coefficients of a plurality of seasonalities, and
identifying a subset of a plurality of seasonalities having respective coefficients that exceed a seasonality threshold.
8 . The method of claim 1 further comprising
measuring, based on said timeseries, a joint distribution of two selected from the group consisting of: a first expected value, a second expected value, a first variance, a second variance, a first moment of at least third order, and a second moment of at least third order.
9 . The method of claim 1 further comprising
calculating, based on a unit root test, a count of differencing operations needed for data stationarity.
10 . One or more computer-readable non-transitory media storing instructions that, when executed by one or more processors, cause:
automatically measuring, for a timeseries, a seasonality and a stationarity differencing count;
assigning the stationarity differencing count to a respective differencing hyperparameter of each machine learning (ML) algorithm from a plurality of ML algorithms;
selecting, based on said assigning, a most accurate ML algorithm from the plurality of ML algorithms;
identifying a values subrange of a seasonality hyperparameter of the most accurate ML algorithm;
automatically selecting, based on said stationarity differencing count, a minimum length of a training subsequence of the timeseries;
tuning, based on the values subrange of the seasonality hyperparameter of the most accurate ML algorithm, hyperparameters of said most accurate ML algorithm; and
training, with a plurality of subsequences of the timeseries that have at least the minimum length, an ML model that is based on; said tuning and the most accurate ML algorithm.
11 . The one or more computer-readable non-transitory media of claim 10 wherein:
the plurality of ML algorithms consists of said most accurate ML algorithm and a plurality of other ML algorithms;
the instructions does not cause tuning hyperparameters of at least one ML algorithm of the plurality of other ML algorithms.
12 . The one or more computer-readable non-transitory media of claim 10 wherein the instructions further cause validating, based on said stationarity differencing count, said most accurate ML algorithm during at least one selected from the group consisting of:
said selecting said most accurate ML algorithm and
said tuning said hyperparameters of said most accurate ML algorithm.
13 . The one or more computer-readable non-transitory media of claim 12 wherein;
said validating said most accurate ML algorithm comprises automatically calculating a forecast horizon duration.
14 . The one or more computer-readable non-transitory media of claim 12 wherein:
said validating said most accurate ML algorithm comprises measuring, for each seasonality of a plurality of seasonalities, a respective fitness of said most accurate ML algorithm;
said fitnesses of said most accurate ML algorithm consists of:
a) a highest fitness for a particular seasonality of said plurality of seasonalities and
b) fitnesses for other seasonalities;
said tuning said hyperparameters of said most accurate ML algorithm is based on said particular seasonality and not said other seasonalities.
15 . The one or more computer-readable non-transitory media of claim 10 wherein the instructions further cause validating the ML model based on said stationarity and at least one selected from the group consisting of seasonality and frequency.
16 . The one or more computer-readable non-transitory media of claim 10 wherein the instructions further cause
measuring seasonality by at least one selected for the group consisting of:
removing linear trends,
fitting a second-degree polynomial regression,
fitting a polynomial regression with an ordinary least squares estimator,
applying an autocorrelation function (ACF) to a differenced timeseries that is based on said timeseries,
calculating a coefficient at specific lagged values of a differenced timeseries that is based on said timeseries,
calculating a coefficient that represent a strength of a linear relationship between respective values in two seasons of a same seasonality,
sorting respective coefficients of a plurality of seasonalities, and
identifying a subset of a plurality of seasonalities having respective coefficients that exceed a seasonality threshold.