Forecast & Optimization

Data Science

Design, simulate and optimise product launches and sales management through a robust approach

Solutions:

Data Science

Technologies:

PyTorch - Python

Overview

With a view to a data-driven transformation, it is becoming increasingly necessary for a company to analyse a data history in order to make important decisions on future strategies and the next steps to implement them.

In any field of application, therefore, having algorithms in production to predict and optimise the problems posed by these challenges becomes strategic for a company: in this way, it is possible to realise the potential capital of its data and to control the impact and performance of its choices. In fact, these algorithms can help to improve the efficiency of production and organisational processes. In particular, the optimisation of production processes has become an increasingly challenging objective for companies in any sector.

Genetic algorithms, embedded in a suitable calculation and results architecture, help to achieve complex prediction and optimisation objectives and to manage a considerable number of possible scenarios, increasing their performance as they come into contact with the data to be learnt.

Challenges

The main problems arising from a ‘traditional’ approach to forecasting and optimisation problems may be the following:

High number of relevant features to be managed, controlled and utilised to achieve the forecast optimum
High calculation time and difficulty in distribution
Complexity in the research and development of the best prediction and optimisation algorithm suited to the customer's particular use case
Difficulty in achieving a satisfactory result to a prediction problem due to the potential number of possible scenarios
Maintenance in production of the developed algorithms complex for the rapidity of modifying the actual scenarios on which the prediction is made

Solution

We designed an innovative approach for the efficiency of prediction and optimisation techniques, processing historical data and applying feature selection and combinatorial techniques borrowed from genetic algorithms

Genetic algorithms are a class of search and optimisation algorithms inspired by Darwin’s theory of evolution. Genetic algorithms use a dataset to find a population of candidate solutions and apply genetic operators such as selection, mutation and outcrossing to generate new candidate solutions. These new candidate solutions are evaluated for their suitability for the problem at hand. The process of generating solutions is repeated until a satisfactory result is achieved.

Genetic algorithms are a class of heuristic search algorithms used to solve optimisation problems. The main reasons for using a genetic algorithm are:

The objective function is not regular (so derivative methods cannot be applied)
A large number of parameters can be a problem for derivative-based methods when there is no definition of the gradient
The objective function is noisy or stochastic
There are multiple local optimum points

Thinking of adopting this category of algorithms leads to several advantages:

Fast search technique: results close to the optimum are produced in reasonable time.
They are suitable for parallelisation of processing.
They are simple to develop

The architecture of the solution designed by Quantyca, thanks to the pliability of the genetic algorithms, can be applied in various areas where optimisation and as-is improvement problems come into play and is particularly suitable in the retail sector, where a very large number of variables and features have to be taken into account.

The overall system designed is articulated through the orchestration of six steps:

a feature engineering process that transforms the data and organises it into an optimal structure on which to train the forecast model
a process of training and controlling the forecasting model of the target variable
a forecasting process for exogenous variables based on seasonality and trends
a configurator of operational scenarios within which to search for the optimal configuration of decision variables
a second configurator to define the sets of decision variables from which to search for optimal combinations
an optimiser, based on genetic combinatory algorithms, which, starting from the information of the operational scenarios, iteratively identifies the best combinations of decision variables