Completing Your Decision Pipeline

Leading a successful data science project is not an easy feat, as you’re typically caught between two worlds.

Business teams have a hard time speaking your language but throw more and more complex problems at your people. Your technical people use an ever-expanding box of magical tools that all seem to solve similar problems. And when the techies are done, you may struggle to correctly attribute the bottom-line results to all the efforts they put in.  

So why would you plan to invest resources in exploring yet another piece of technology? And a decision intelligence technology, at that? 

Plan for Impact

Data science often starts with data and tries to spot useful patterns. The implicit assumption is that properly understanding (modeling) the data would somehow ultimately translate to business benefits.  

But what if you could start with the end goal: achieving your business objectives?   

Mathematical optimization—a powerful decision intelligence technology—explicitly describes the business objectives as part of the model. After adding the freedoms and constraints, the role of optimization is to efficiently find recommendations that lead to optimal results for those objectives. 

Plan for Complexity

While the ability to identify the best way to achieve your business objectives may sound attractive, it’s typically not the objectives that make the problem difficult. Many business rules interact with each other and together determine how your business can act, based on the reality your machine learning models predict.  

Your team may start with simple, heuristic approaches to translating predictions into plans. As your business users add complexity iteratively, however, you will soon find that additional rules either cannot be handled at all or prevent you from finding an optimal plan.  

Mathematical optimization separates the problem definition from the algorithmic toolbox. Your team focuses on describing the problem with all its complexities, while the optimization software takes care of finding the optimal solution to that problem. 

Plan for Uncertainty

Assuming your data science projects have led to a perfect forecast, mathematical optimization would complete your decision pipeline and give optimal decisions based on your data.  

But the reason your team spends so much time on those data science models is that reality is very hard to predict. At some point, accuracy cannot be improved any further, while your business users still ask, “What if?”  

The answer is to generate more than “just” the optimal plan for a single future. With optimization technology, you can perform sensitivity analysis and understand how changes in input data (and forecast) would change the cost of your optimal plan. Or, you could run a wide range of scenarios and compare optimal plans for each scenario to better understand how optimal decisions might change with your data.  

Plan for Change

So far, we’ve looked at all the things that can be done before decisions are executed by your operational teams. At some point, one of the many scenarios you considered becomes reality. The truth is, reality is often slightly different from every scenario you have considered.  

So how do you adapt your plan now?  

Making many manual adjustments to a carefully optimized plan typically defeats the purpose of optimization. But there’s often no time to completely rebuild your plan—not just because of computation time, but because distributing a completely new plan to all stakeholders is practically impossible.  

Instead, you’ll want to balance minimizing changes to the original plan against your original objectives to generate a new plan within seconds or minutes. Mathematical optimization solvers let you do just that.  

Plan for Success

Finally, after your plan has been successfully adjusted and executed, without any doubt your data science team deserves all the credit. But typical machine learning use cases make it difficult to attribute success (in terms of business KPIs) to the quality of the trained model. What part of a profit increase can be linked to the accuracy of your model—and how would you explain the causality?  

Mathematical optimization has an interesting answer to this challenge. Take your existing software or most experienced planner, give them one or a few sets of input data, and let them create the best plan possible following the same rules as your mathematical optimization model. Then, compare the results by the objectives you agreed upon. Conversations about success attribution will suddenly become a lot easier! 

Start Your Optimization Journey

Even with a perfectly optimized plan, the quality of input data, forecast accuracy, and the level of operational excellence will impact your bottom-line results. But by adding mathematical optimization capabilities to your toolbox, you can properly address impact, complexity, uncertainty, change, and, ultimately, success. Sound like a plan? 

If you’re ready to learn more, check out our free learning resources for data scientists and business leaders.