The basic steps of Techaisle's Predictive Modelling are - Data Integration, Non-linear modelling, and ANN embedding.
1. Data Integration
Most customers have large amounts of transaction data. This data shows what is happening. However, for effectively managing the business the customer needs to know the 'Why' behind the transactions. Techaisle answers the 'Why' question through primary research, integrate these findings with the transactions. This process of integration is as much an art as science. It needs considerable experience and domain knowledge to achieve an effective integrated data base.
2. Non-linear modelling
Response to a business decision is non-linear. Linear models have been found to be ineffective in building business decision models. Techaisle's modelling is primarily non-linear and hence is more realistic. We routinely use polynomial, exponential and mixed models for the response variables.
3. Artificial Neural Networks
Techaisle uses a two pronged approach to modelling. The data analysis is carried out using the statistical tools and analytics is carried out using Artificial Neural Networks (ANN). ANN offers an enormous range of non-linear models that are stable over a period of time and hence ideal for predictive analytics.
Objectives of Predictive Analytics
Techaisle's objective is to help improve business performance through reliable predictive analytics. Techaisle's predictive modeling framework includes: Integration, Simulation, and Optimization.
Unique features of Techaisle's predictive analytics services are:
- Minimal usage of historical data: especially with ever-changing technological landscape, historical data is of little use unless it is for a mature technology area
- Accommodate large number of variables: in typical linear models the number of variables become a huge limitation
- Add new variables on-the-go: New variables can introduced into the model at any time. For example, if a new PC form-factor is introduced, ANN can easily incorporate it into the already existing model which is difficult in statistical modeling.
- Create forecasts prior to product introduction: and calibrate after 1-2 weeks of sales; this also allows to measure the impact of competition and price changes
Techaisle's predictive modeling analytics team successfully overcomes the limitations of existing modeling techniques (multiple regression, logistic regression, structural equation model, regression trees, churn models) used by most marketers:
- Almost all models need a large number of measurements
- Error is determined by the model, for example, if R2 is 66 percent, it cannot be programmed to achieve 90 percent accuracy. However, variables can be included or excluded to achieve higher accuracy but perhaps variables excluded could be very essential
- Accuracy and predictability have to be balanced
- Modeling is expensive
- Models assume that business is passive, but businesses are not based on what has happened
- Cannot utilize hidden layers, for example, if a function has 70 percent fit and the marketer wants to improve the fit existing models are unable to handle changes
Techaisle's predictive analytics team has also successfully solved the market-mix problem that faces most marketers to understand the relationship between ATL/BTL and sales. Marketers are seeking 90 percent accuracy levels but most existing models have effectively not gone beyond 80 percent accuracy level. Techaisle's proven solutions based on Embedded Artificial Neural Networks (ANN) consistently deliver 90%+ accuracy levels.
Embedded Artificial Neural Networks
ANNs are extremely powerful and in experienced hands it can solve most complex problems that statistical techniques cannot solve with a high degree of accuracy. Statistical techniques are useful but have inherent disadvantages. Quite a few of these limitations can be overcome by ANNs.
Innovative Data Integration
To make ANNs deliver a high accuracy it is necessary to get different data together on a single usable data structure. This has been achieved through a novel Integration method. Because of this innovative file structures it has been possible to embed a series of ANN modules in to the solution. Techaisle has successfully integrated economic, environmental, market research and transaction data in its solutions
Short Historical Data
Techaisle's solutions are designed to use short historical data. Most statistical models require data of 60-100 data points, weekly or monthly, to develop a model. In many categories historical data of five or ten years is neither available nor relevant since market environment changes fast.
Techaisle solutions have simulation algorithms embedded in the model which can be used to perform a number of 'What if?' scenarios.
Optimizer is another feature that can be built and is optional. Using the optimizer in conjunction with the simulator increases the flexibility geometrically.
Broad Range of Application Areas
The solutions are capable of addressing a wide range of problems in brand response, retail sales management, pricing, cannibalization etc.