Today’s focus on generative AI provides a unique opportunity for organizations to leverage tools housing large and complex data sets.  Due to the nature of the technology, you can train models to provide much greater return on your investment when you have a lot of data to provide feedback to your models. Organizations using Planning Analytics are uniquely positioned to take full advantage of this opportunity in time.

The nature of the TM1 Database engine as a functional database, in-memory technology, provides an exceptional foundation for someone to build up a predictive solution.  This is due to the built-in business logic that is inherently embedded within the results of the TM1/PA solution.  For example, a company that is using Planning Analytics for revenue planning will most likely include not just their top line revenue numbers, but also the various drivers such as Price, units, discounts, etc. to achieve accurate gross through net forecasts.  This driver-based planning model would cross all of the important dimensions of their data both from a reporting and management perspective. 

There are a variety of ways that a common model such as the one above can be applied to supporting a modern AI based data science approach:

Linear Regression and Time Series Forecasting

One of the most used data science approaches in the planning world is the time series forecasting methodology.  This is kind of the tried-and-true method and has been around for many years.  This approach adheres to the ethos “The Past will Predict the Future” and requires the use of historic actual data to project what is going to happen in the future.  This method is sometimes referred to as trending, which is related, but it uses much more complicated statistical principles to achieve a result.  One of the key aspects of Linear regression time series forecasting is the generation of confidence or error bands.  Due to this being a mathematical projection, you will have statistically significant errors built into the model based on how consistent your historical data is. 

For example, a historical sales number is likely going to be impacted significantly by market conditions and pricing sensitivity due to macro-economic conditions outside of your control.  Because of this we would expect the sales forecast to have a large window of both upside and downside particular as our prediction window goes longer.  Conversely, things like fixed expenses such as insurance, or utilities are likely to have much smaller windows of error due to consistent cost structure to those expenses through time.  Even with something like utilities that may vary significantly from season to season, the Time Series model can take that into account and generate a seasonally weighted projection with a much lower error rate. 

For forecasts that are consistent with a long history of data this method is quick, easy, and accurate.

Decision Optimization

When we talk about decision optimization, we are really talking about the ability to generate a mathematically balanced result for some problem statement, based on aligned constraints.  By doing this we can develop high performing forecasts and plans that are more achievable and potentially more accurate due to the inclusion of restrictions in things like production volume, supply chain capacity, or other limitations that are not as flexible as we may like.  A great example of this is Inventory modeling for a retail type of organization.  One of the key measures of success for retail is the elimination of “Out of Stock”, thus eliminating lost sales simply because the SKUs are not available.  In a decision optimization model, we can pull in our demand plan by location, inventory space by location, and current inventory positions and run them through the decision optimization.  The mathematical model behind this will take all the constraints we have and will generate the “Best Fit” inventory position to achieve our provided demand plan.

Back to our Driver Based planning model, one common question is, how will we achieve our revenue targets?  By leveraging a decision optimization approach, we can quickly and efficiently generate various solutions by leveraging the existing drivers we have defined within our TM1 model, thereby decreasing our time to iterate and review and increasing our ability to action on the results and develop a proper plan of action.  What previously would require weeks or months of data entry and analysis can now be done within days or hours.

When you have mixes of data with confidence in one or more of the values, decision optimizations would be a great fit to apply to your model.

Generative AI

There has been a lot of news lately around generative AI.  This is a cutting-edge technology made possible by advancements in computing technology and data availability.  It is a complex field with multiple ways in which it can be applied.  For the functional database user this commonly takes the form of what’s known as a Deep Learning model.

Deep learning is a model that attempts to replicate how a human would approach understanding and evaluating what might be happening with a data set or data model. It takes a data model with a defined goal at the end (e.g. variance from actuals) and is set up to score itself based on guessing the parameters of the model and evaluating how close it gets to the goal.  The model is run a significant number of times simultaneously, sometimes thousands of times, and the best performing models are then used as the starting point for the next “generation” of models.  This cycle of guessing, checking, and eliminating the worst performers continues until there’s no further increase in score. 

There are a variety of advantages to this method, including the potential for significantly more accurate results.  You can continue to train the model in a continuous improvement approach by leveraging newer data as it becomes available.  This is an excellent approach for any situation where you don’t know if your drivers are correct, or they aren’t tracked in your source system.

Conclusion

There’s no one size fits all approach when it comes to the data science and AI tool kits, but there are benefits to be had within your TM1 data model.  By leveraging the modern approached developed for everything from the Chat GPTs and Google Gemini to some of the biggest companies in the world such as Tesla and Amazon, you can achieve more accurate, consistent, and data driven results quickly and easily allowing you to focus more on how to prevent the downside and increase the upside.