Sas Institute Inc.

Mahesh V Joshi

Mark Roland Little

Jan Chvosta

Michele Angelo Trovero

Richard Potter

Michael James Leonard

Iman Vasheghani Farahani

Rajib Nath

Phillip Mark Helmkamp

Dmitry V Golovashkin

Richard Patrick Fahey

Youngjin Park

Yung-Hsin Chien

Javier Delgado

Vilochan Suresh Muley

Ann Mary McGuirk

Steven Christopher Mills

Hongtao Hu

Ran Bi

Kritika Misra

Sounak Paul

A system and method include tuning hyperparameters for an ARIMA model using a derivative free approach by determining a set of initial hyperparameter values, fitting an ARIMA model to the set of initial hyperparameter values, selecting a tuning method for the set of hyperparameters, responsive to selecting a single-objective method, computing a first objective function value from time-series data applied to the ARIMA model based on the set of initial hyperparameter values, or responsive to selecting a multi-objective method, computing at least a second objective function value and a third objective function value from the time-series data applied to the ARIMA model based on the set of initial hyperparameter values, determining whether a stopping criterion for tuning the set of hyperparameters has reached, responsive to determining that the stopping criteria has reached, outputting a set of tuned hyperparameter values.

Hyperparameter tuning in autoregressive integrated moving average (ARIMA) models

A system described herein can generate a graphical user interface (GUI) for a piece of forecasting software. The GUI can include graphical nodes arranged on a drag-and-drop canvas to define an overall forecasting pipeline. The graphical nodes can include a hierarchical modeling node that enables a user to define a time series hierarchy. The hierarchical modeling node can also enable separate level pipelines to be customized for each level of the time series hierarchy. The level pipelines can form subparts of the overall forecasting pipeline. The system can then execute the level pipelines to generate multiple forecasts, where each forecast corresponds to a respective level of the time series hierarchy. In some examples, the system can execute a reconciliation process on the forecasts to generate reconciled forecasts. Each reconciled forecast can correspond to one of the forecasts. The system may then generate one or more visualizations of the reconciled forecasts.

Hierarchical modeling node for visual forecasting

A new value is written from a dataset to a data structure comprising a set of sorted values. The new value replaces an oldest value and is inserted in a sorted position. The data structure is modified by subtracting a median value from each value of the set of sorted values to obtain sorted signed deviation values. The sorted signed deviation values are segmented to obtain data substructures comprising subsets of sorted absolute deviation values. A binary search is performed on the data substructures to identify a median absolute deviation value. A difference is computed between a particular value and the median value, and based on whether the difference is less than a threshold value computed from the median absolute deviation value, an outlier decision output is generated indicative of whether the particular value comprises an outlier value.

Optimized hampel filtering for outlier detection

In one example, a computer system can generate a graphical user interface (GUI) for forecasting software including a drag-and-drop canvas with a set of rearrangeable nodes defining a forecasting pipeline. The computer system can detect a user interaction for attaching an external-language execution node to the pipeline, which can be used to insert custom code defined using an external programming language. The computer system can receive the custom code. The computer system can receive a user input to initiate execution of the pipeline. The computer system can generate wrapped custom code by augmenting the custom code with additional program code including shared variables. The computer system can provide the wrapped custom code to a set of execution threads configured to execute the wrapped custom code as part of the pipeline to generate one or more forecasts. The computer system can output the forecasts in the GUI.

External-language execution node for visual forecasting

A computing device provides distributed estimation of an empirical distribution function. A boundary cumulative distribution function (CDF) value is defined at a start of each region of a plurality of regions. An accuracy value is defined for each region. (a) First equal proportion bins are computed for a first sample of a first marginal variable using the defined boundary CDF value for each region. (b) Second equal proportion bins are computed for the first sample of the first marginal variable within each region based on the defined accuracy value for each region. (c) The computed second equal proportion bins are added as an empirical distribution function (EDF) for the first marginal variable. (d) (a) to (c) are repeated for each remaining sample of the first marginal variable. (e) (a) to (d) are repeated with each remaining marginal variable of a plurality of marginal variables as the first marginal variable.

Two-stage distributed estimation system

Metadata received from each worker computing device describes EDF estimates for samples of marginal variables stored on each respective worker computing device. Combinations of the EDF estimates are enumerated and assigned to each worker computing device based on the metadata. A request to compute outcome expectation measure values for an outcome expectation measure is initiated to each worker computing device based on the assigned combinations. The outcome expectation measure values computed by each worker computing device are received from each respective worker computing device. The received outcome expectation measure values are accumulated for the outcome expectation measure. A mean value and a standard deviation value are computed for the outcome expectation measure from the accumulated, received outcome expectation measure values. The computed mean and standard deviation values for the outcome expectation measure are output to represent an expected outcome based on the marginal variables.

Techniques for estimated compound probability distribution are described herein. Embodiments may include receiving a compound model specification comprising a frequency model and a severity model, the compound model specification including a model error comprising a frequency model error and a severity model error, and determining a number of frequency models and severity models to generate based on the received number of models to generate. Embodiments include generating a plurality of frequency models through perturbation of the frequency model according to the frequency model error, and generating a plurality of severity models through perturbation of the severity model according to the severity model error. Further, embodiments include dividing generation of a plurality of compound model samples among a plurality of distributed worker nodes, and receiving the plurality of compound model samples from the distributed worker nodes, and generating aggregate statistics from the plurality of compound model samples.

Techniques for estimating compound probability distribution by simulating large empirical samples with scalable parallel and distributed processing

Techniques for estimated compound probability distribution are described. An apparatus may comprise a configuration component, perturbation component, sample generation controller, an aggregation component, a distribution fitting component, and statistics generation component. The configuration component may be operative to receive a compound model specification and candidate distribution definition. The perturbation component may be operative to generate a plurality of models from the compound model specification. The sample generation controller may be operative to initiate the generation of a plurality of compound model samples from each of the plurality of models. The distribution fitting component may generate parameter values for the candidate distribution definition based on the compound model samples. The statistics generation component may generate approximated aggregate statistics. Other embodiments are described and claimed.

Techniques for compressing a large distributed empirical sample of a compound probability distribution into an approximate parametric distribution with scalable parallel processing

Techniques for estimated compound probability distribution are described herein. Embodiments may include receiving, at a master node of a distributed system, a compound model specification comprising frequency models, severity models, and one or more adjustment functions, wherein at least one model of the frequency models and the severity models depend on one or more regressor and distributing the compound model specification to worker nodes of the distributed system, each of the worker nodes to at least generate a portion of samples for use in predicting compound distribution model estimates. Embodiments may also include predicting the compound distribution model estimates based on the sample portions of aggregate values and adjusted aggregate values.

Techniques for estimated compound probability distribution are described. An apparatus comprising a configuration component, perturbation component, sample generation controller, an aggregation component, a distribution fitting component, and statistics generation component. The configuration component operative to receive a compound model specification and candidate distribution definition. The perturbation component operative to generate a plurality of models from the compound model specification. The sample generation controller operative to initiate the generation of a plurality of compound model samples from each of the plurality of models. The distribution fitting component to generate parameter values for the candidate distribution definition based on the compound model samples. The statistics generation component to generate approximated aggregate statistics.

Computer-implemented systems and methods provide paginated search results from a multi-table database that stores hierarchically arranged data nodes. A query request is received that identifies one or more constraints and one or more monitoring attributes. Records of the multi-table database are filtered to create a view containing only records which meet the constraints. Monitoring attributes associated with a first n records of the view are accessed to generate an output view, where n is the number of records that may be displayed on a single page, and the output view is displayed on a user-viewable medium.

Computer-implemented systems and methods for providing paginated search results from a database

Systems and methods for reconciling a forecast are presented. A method can be used that receives a plurality of hierarchical forecast data sets. An output child data set including an index value and a status indicator representing an unprocessed state is generated. A particular parent data set forecast is identified from a parent data set. Locations for a group of one or more child data set forecasts that are children of the particular parent data set forecast are identified and accessed. A reconciliation operation is performed, a particular child data set forecast is adjusted and stored in a record, and a status indicator for the record is modified.

Computer-implemented systems and methods for forecast reconciliation

Systems and methods for reconciling a forecast for a dimension based upon data that is associated with the dimension. A method can be used that includes generating a plurality of forecasts for the dimensions such that the forecast of a first dimension is generated independently of a forecast of a second dimension. The forecast of the first dimension has a constraint that is influenced by the forecast of the second dimension. A reconciliation is performed between the forecast of the first dimension and the forecast of the second dimension in order to determine how the constraint of the first dimension's forecast is to influence the first dimension's forecast.

Systems and methods for providing estimations for a product for purchase at a plurality of stores. Groups of stores are generated based upon similarity of store demand data. For each group, a distribution is determined with respect to the attribute of the product. The distribution is used to provide estimations with respect to the product to be provided at the stores.

Growing community of inventors

Cary, NC, United States of America

Mahesh V Joshi