Year: 2024

Demand Forecast — A Value-Driven Approach with 5 Insights

The end goal of forecast is not accuracy, but tangible value creation. Let’s explore 5 insights to maximise the value of demand forecasts

Intro

Hey there! January is the perfect time for planning and making a big impact. As a data scientist, you’re often asked to build forecast models, and you may believe that accuracy is always the golden standard. However, there’s a twist: the real magic lies not just in accuracy but in understanding the bigger picture and focusing on value and impact. Let’s uncover these important aspects together.

1. Value is the ultimate goal of demand forecast

Regarding forecasts, we should first align on one thing: our ultimate goal is about creating real value. Real value can manifest as tangible financial benefits, such as cost reductions and revenue increases, or as time and resources that you free up from a forecast process. There are many pathways which start from demand forecast and end in value creation. Forecast accuracy is like our trusty compass that helps us navigate toward the goal, but it’s not the treasure we’re hunting for.

Your to-dos as a data scientist:

• Discuss with your manager and team the purpose of the demand forecast. Is its goal to set accurate sales targets? To lower inventory levels? What are the underlying concerns behind these forecast numbers?
• Create a simple business case to translate forecast accuracy metrics (bias, MAPE) into financial terms. If this task seems daunting, ask for help from your friends on the business side. Together, you will learn A LOT about the business and the value of your demand forecast.
• Evaluate your business case to identify the most crucial aspect of the forecasting exercise. Is it reducing bias (particularly over-forecasting) to decrease inventory levels? Is it assessing the impact of discounts on various products categories (which might be better served by an elasticity model)? Or is it more about lowering MAPE to prevent the supply team from constantly reacting to unpredictable purchase orders in a crisis

By clearly connecting the dots between forecasting elements and their value, you’ll feel more confident about where to direct your energy and brainpower in this forecasting exercise.

2. Small tweaks in the process can go a long way

In forecasts, you can add value in two areas: process and model. As data scientists, we may be hyper-focused on the model, however sometimes, a small tweak in the process can go a long way. The process that produces the forecast can determine its quality, usually in a negative way. Meanwhile, the process that begins with the forecast is the pathway leading to value creation. Without a good process, it would be hard for even the best model to create any value.

Your to-dos as a data scientist:

• Learn about the “best practices” in forecasting. This can be tricky since different industries and business models have their own definitions of what “best practices” are. But some principles are universally valid. For instance, forecasts should be generated automatically on a regular basis; manual overrides should be rare and only for solid reasons; and forecasts ought to trigger clear decisions and actions like preparing production, adjusting inventory, or ramping up promotions.
• Check out these “best practices” and see if you’ve covered all your bases. If yes, awesome! You’re ready for the next challenge. If not, dig a bit deeper. Ask yourself who or what is holding things back. What are the smallest changes that could improve the whole forecasting process? I’d really recommend grabbing a coffee with a key player in this area. You might be surprised at the impact you can have by swaying just one person in the forecast process.

3. Clearer process mapping enables better model-building

Even when the process is too ingrained to change, having a clear understanding of the process is still tremendously valuable. It allows you to focus on the key features that are most pertinent in the chain of decisions & actions.

For instance, if production plans need to be finalised two weeks in advance, there’s no need to focus on forecasts for the upcoming week. Likewise, if key decisions are made at the product family level, then it would be a waste of time to look at the accuracy at the individual product level. Let the (unchangeable) process details define the boundaries for your modelling, saving you from the futile task of boiling the ocean.

Your to-dos as a data scientist:

• Pair up with a business-savvy buddy and sketch out a diagram of the forecasting process. Make sure each step includes these elements: the decision being made, the inputs for the decision, who’s making the decision, and the outcomes that follow. Remember, this isn’t an easy task and we’re not aiming for perfection. Gather as much info as you can and piece it together on paper.
• Next, take a look at your diagram (which might look a bit overwhelming with all its circles and such) and try to pinpoint the most critical decisions in the entire process. Figure out what kind of forecast is essential for making solid decisions at these points: do you need a 6-month forecast at the product family level, or a weekly forecast for each specific product package variant? These are the crucial issues that your top-notch modelling skills and data science knowledge will tackle.

4. Improve explainability for better alignment and adoption

On the modelling side, explainability should be a top priority, as it significantly enhances the adoption of the forecasts. Since our ultimate goal is value creation, forecasts must be used in business operations to generate tangible value.

This could involve using them in promotion planning to increase revenue or in setting inventory targets to reduce stock levels. People often have the choice to trust or distrust the forecast in their daily tasks. (Ever been in a meeting where the forecast is dismissed because no one understands the numbers?) Without trust, there is no adoption of the forecast, and consequently, little value can be created.

On the contrary, when the forecast numbers come with an intuitive explanation, people are more likely to trust and use them. As a result, the value of an accurate forecast can be realised in their daily tasks and decisions.

Your to-dos as a data scientist:

• Think about the forecasting process and consider whether people want and need a better understanding of your forecast model. I’d say if the forecast is used for humans to make medium or long-term decisions (like budgeting, pricing, or capacity planning), explaining it is crucial to build trust in the data and prompt a decision.
• You also need to grasp how decision-makers intuitively interpret or anticipate forecast numbers. Then, tailor your explanation to speak their language. This is the tricky part — you’ll have to rework your feature importance, Shap values, and regression coefficients into terms like “the impact of a 1% price increase.” Don’t hesitate to ask your business-savvy friend for help and test your explanation on them to see if it makes sense

5. Simulate scenarios to facilitate decision-making

Scenario simulation naturally extends from explainability. While an explainable model helps you understand forecasts based on anticipated key drivers (for example, a 10% price increase), scenario simulation enables you to explore and assess various alternatives of these anticipations or plans. You can evaluate the risks and benefits of each option. This approach is incredibly powerful in strategic decision-making.

So, if you’re tasked with creating a forecast to determine next year’s promotion budget, it’s crucial to align with stakeholders on the key drivers you want to explore (such as discount levels, packaging format, timing, etc.) and the potential scenarios. Build your forecast around these key drivers to ensure not only accuracy, but also that the model’s explanations and scenarios “make sense”. This might mean anticipating an increase in demand when prices drop or as holidays approach. But of course, you need to figure out, together with the key stakeholders, about what “make sense” really means in your business.

Your to-dos as a data scientist:

• Chat with people who make decisions to figure out which hypothetical scenarios they want to be prepared for. Have them identify key factors and set the scene: a 10% inflation spike, supply disruption of a crucial raw material, a natural disaster, and so on. Ask them to rank these scenarios and factors in order of importance, so that you can prioritise.
• Next, see how your forecast model stacks up. Try to create simulated forecasts for some of these scenarios and factors, starting always with the most important ones.
• Check with your business-savvy friend to ensure your simulations are realistic. You might need a few tries to tweak your models and get everything just right. Like with explanations, using business language to narrate the story is key in this task. Don’t hesitate to ask for help. It’s a learning opportunity for both you and whoever assists you.

Wrapping up

Alright, I know this seems like a lot to take in. You might be thinking, “So, in addition to crunching data and training models, do I also need to delve into process analysis, come up with an explanatory model, and even build a simulation engine for forecasting?”

No need to worry, that’s not exactly what’s expected. Look at the bigger picture, will help you pinpoint the key aspects for your forecasting model, figure out the best way to build them, and connect with the right people to enhance the value of your forecast. Sure, you’ll have to add a few extra tasks to your usual routine of data crunching and model tuning, but I promise it’ll be a rewarding experience — plus, you’ll get to make some business-savvy friends along the way!

If you want to go deeper than this simple framework, I have also compiled a comprehensive list of questions in this article to cover all aspects related to demand forecast. Have fun with your forecast project and maximise your impact on the world!

Demand forecast — a value-driven approach with 5 key insights was originally published in Towards Data Science on Medium, where people are continuing the conversation by highlighting and responding to this story.