Using this kind of approach allows us to incorporate business know-how naturally in the form of priors, explore the space of models, and include uncertainty to find the best distribution of models, resulting in a more robust solution.
State-of-Art Bayesian Heuristic Methods
Using MCMC, SGD and Variational training methods enables automated training with initialization, sampling and stopping criteria.
Transport Gaussian Process
This kind of non-parametric process allows accurate modeling of the base sales, revealing structure in seemingly random behavior.
Proprietary Algorithm Implementations
Our in-house built software library gives us full control and understanding over the algorithms and methods we implement, using open source libraries for low-level operations only.
Models are trained in stages, allowing us to keep focus on individual aspects one at a time. This also minimizes impact and reprocessing costs due to changes in scope or data.
All used data is transformed to a standard scale, creating a common modeling parameter space across all projects.