Autoregressive Tabular Forecasting for Multi-Seasonal Time Series
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing forecasting algorithms struggle with accurately predicting future values in timeseries data that exhibit complex temporal patterns and multiple seasonalities, especially when influenced by exogenous variables, and often result in inaccurate predictions due to out-of-bound issues and the assumption of identical statistical distributions between training and testing datasets.
Innovation Solution
A one-shot, automated approach that utilizes autoregression to discover multiple seasonalities and generate synthetic features, coupled with isolated modeling of exogenous variables, to enhance forecasting accuracy by capturing autoregressive correlations and handling trend compensation effectively.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If state of the art statistical modelling formats are used for forecasting, then the forecasting process is simple to implement, but the accuracy is poor due to inability to capture complex autoregressive correlations and multiple seasonal patterns
Solution Approach 1:
The patent segments the forecasting task into multiple components: (1) detecting multiple seasonal patterns separately using autocorrelation analysis, (2) modeling each seasonal component with dedicated lag features, (3) handling exogenous variables independently, and (4) combining results through additive or multiplicative seasonal decomposition. This segmentation allows the model to capture complex patterns while maintaining manageable complexity through modular feature engineering.
Solution Approach 2:
The patent transforms the time series forecasting problem from a univariate regression task into a multivariate problem by introducing multiple seasonal lag dimensions. Specifically, it creates separate lag features for different seasonal periods (e.g., daily, weekly, monthly patterns) and combines them in a multi-dimensional feature space, allowing the model to capture patterns across multiple temporal dimensions simultaneously.
2Measurement precision
If future values of exogenous variables are assumed to be available for forecasting, then the model can utilize these variables as input features, but this assumption is often invalid in practice limiting model accuracy
Solution Approach 1:
The patent performs preliminary action by forecasting exogenous variable values in advance using their own historical patterns and autocorrelations before using them as inputs for the primary forecasting model. This allows the system to handle cases where future exogenous values are unknown, as the model first predicts these values and then uses them for the main forecasting task, eliminating the need to assume future exogenous values are available.
Solution Approach 2:
The patent introduces an intermediary forecasting step that bridges the gap between historical data and future predictions. Instead of directly using unavailable future exogenous values, the system employs an intermediate model to generate predicted exogenous values, which then serve as inputs to the primary forecasting model. This intermediary approach allows the system to handle unknown future exogenous values while still leveraging their potential predictive power.
3Ease of operation
If training and testing datasets are presumed to have the same statistical distribution in generic regression, then the standard regression framework can be applied, but this presumption fails in forecasting problems with trends causing out-of-bound predictions
Solution Approach 1:
The patent segments the forecasting problem into detrending and forecasting components. It first removes trends and seasonal patterns from the time series to create a stationary component, then applies regression modeling only to this stationary part. This segmentation allows the use of standard regression frameworks while avoiding the out-of-bound prediction problem caused by trends, as the model learns patterns from the stationary component rather than trying to predict trending values directly.
Solution Approach 2:
The patent transforms the non-stationary time series data into stationary data by applying parameter changes through differencing and detrending operations. It modifies the statistical properties of the input data to remove trends and stabilize mean and variance, thereby making the training and testing datasets have comparable statistical distributions. This allows standard regression frameworks to be applied reliably without producing out-of-bound predictions.
Data Source
AI summary
For machine learning of timeseries forecasting, here is discovery of multiple seasonalities by autoregression. A multivariate timeseries contains a variable that has a first seasonality that has a first period and a second seasonality that has a second period that is longer than the first period. Many local maxima of an autocorrelation of the variable in the timeseries are selected. For each local maximum as a distinct lag, a candidate feature that lags the variable based on the distinct lag is inserted into the timeseries. Feature selection selects a minimal subset of the candidate features. Based on the minimal subset of features, a feature vector that represents a point in the timeseries is generated. From the feature vector, a future value for the variable is predicted.