Demand forecasting is evolving. The old ways are dead: if you’re still doing human-based forecasting, you’re behind the curve. The future of forecasting leverages AI and advanced analytics to optimize inventory levels for dramatic cost savings.
Tammy Waggoner and Bennjamin Myers have lived this shift, and on May 21, they shared their insights at Nousot’s inaugural Data Connect, a free online forum to explore the latest innovations and pain points in the data world. With their experience in CPG and manufacturing organizations, Tammy and Benn shared lessons learned, real-world use cases, and business impact. Here are five key questions that they answered during the session.
1. Why invest in demand forecasting?
Demand forecasting really breaks down to four steps: 1) generate your statistical forecast, 2) provide external adjustments (such as an economic forecast for a particular industry), 3) provide strategic adjustment (internal changes such as you’ve grown your sales team and can expect more sales), 4) finalize the consensus forecast you build alignment around.
Once you’ve taken these standardized steps, there are three areas where you can see potentially massive payoffs.
Save Costs
You can decrease your holding costs and improve inventory turns. This lets the business run more efficiently and have more cash flow. It also typically translates into very tangible ROI.
Maximize Opportunities
You can find opportunities across your supply chain to maximize returns. On the customer side, this could be avoiding out of stock events. On the supplier side, this could be creating better fulfillment relationships up and down the supply chain.
Add Automation
By adding automation, you can give time back to your demand planners. Demand forecasting can take up a big chunk of time across a large swath of stakeholders. You don’t want to pull everyone out of the loop, but many of those tasks can be automated and streamlined.
2. But why is demand forecasting such a tough problem to solve?
Demand forecasting solutions are plagued by three common problems
Challenging series to model
Many data series can be very challenging to model. For example, there may be erratic periods where nothing is sold, and it’s difficult to predict sales. If you’re trying to forecast and plot errors, the error rate could look very high.
Difficult to scale
Forecasting can also be tough to scale. To put a great model in the hands of the people who need to use it at the time they need it often just doesn’t scale because of the time it takes to turn around the forecast.
Painful to integrate
Data scientists can make great forecasts, but demand planners have to implement it. A great model in a notebook that is then tough to implement means it won’t be used. The handoff point is where a lot of issues can happen, leading to models sitting on shelves never getting used.
3. What’s the best approach to take to demand forecasting?
There are three main approaches to demand forecasting: statistical forecast, machine learning algorithms, and deep learning.
Statistical forecast
On the plus side, statistical forecasting is typically simple, consistent, fast, and easy to modify. However, the downsides are that statistical forecasting can miss more complicated trends and requires you to manually account for external factors.
Machine Learning Algorithms
Machine learning algorithms are capable of driving a high degree of accuracy and are highly customizable. On the other hand, this approach struggles to process time series and requires a lot of feature enhancements to accommodate. Machine learning algorithms also take a lot of attention to tune and maintain.
Deep learning
Deep learning sits in between the other two approaches. It can drive high-performing results at scale. On the flip side, deep learning models are tough to tune and require a certain scale of data. Deep learning does not work well if you only have a moderate amount of data.
So what is the best approach? The truth is there is no silver bullet, and it really depends on your data.
4. What’s the importance of data profiling?
Data profiling is a key way to ensure the accuracy of your forecast. Benn said, “So much of good data science is doing good data engineering. And forecasting is no different.”
Choosing the right approach or combination of approaches comes down to: What does the data you’re trying to predict look like? Data profiling is bucketing your data into profiles, and different profiles require different techniques.
Common Data Profiles
Four series types that are most important to account for are:
Erratic
- High volume, high variability
- A lot of opportunity to get it right or wrong
Intermittent
- Slow moving products with low variance
- Really difficult but aren’t usually the major drivers of your business
Lumpy
- Low volume, high variance
- Also typically not major drivers of your business
Smooth
- High volume, low variance
- Always selling in a more routine way which is easiest to forecast
- Also the ones that are harder to drive significant improvement in the forecast
It’s important to think about and treat each of these data profiles differently if you want to ensure you’re building around all the factors. However, many companies tend to apply one blanket approach, leaving a lot of value on the table.
5. Why do forecasting efforts often fail?
Not embracing an analytics mindset
Forecasting is very measurable, but you have to ensure you’re measuring the right thing. If you are measuring properly, it’s also important to create an environment where you’re trusting the process and metrics and not trying to treat every edge case. Instead, focus on improving the overall process – not every series.
Focusing on the wrong metrics
When thinking about wrong metrics, people tend to focus on MAPE or bias. But ultimately, you need to ensure that your metrics reflect what you’re trying to accomplish. Forecasts can perform poorly because they’re focusing on the wrong things. Align your metrics to your business goals.
Building tech not business solutions
Many organizations are set up with a hub and spoke model for IT, building capabilities that get pushed out to particular business units. Demand forecasting has so much more nuance and has to fit into a lot of different systems. It’s also really difficult to drive adoption, and more importantly, trust in this area. Taking an embedded approach and driving solutions through demand planners ensures alignment, trust, and increased adoption.
Giving up too early
Many organizations don’t see immediate returns and give up too early. As Benn said, “It’s a cliche but it’s true – it’s a journey. When you do something new like this, there’s going to be some toe stubbing. You’ll try some models that may look cartoonish, but the key is to iterate quickly and improve the forecast across the board.”
If organizations stick with it (and we’re talking weeks and months rather than years), organizational buy-in can turn into 20-30% accuracy lifts.
Tammy compared it to big ERP implementations: “I bet some of you have been through an ERP implementation and you don’t give up when that doesn’t go well (and I haven’t been through one that did go well). So you keep pushing forward to get it right. You don’t throw out the ERP baby with the bathwater. Demand forecasting should be approached the same way.”
We’ve worked with countless organizations to realize these accuracy gains and would welcome conversations on how you’re approaching demand forecasting. Reach out to Tammy and Benn to schedule a conversation.
If you’d like to see the full recording of the event, watch it here:
* This content was originally published on Nousot.com. Nousot and Lovelytics merged in April 2025.
