As stewards of the forecasts in a CPG organization, demand planners face daily struggles as they strive to determine and forecast product demand with accuracy. And while demand forecasting continues to be a problem, it’s just one piece of the puzzle. In addition to accuracy, demand planners need the agility and ability to proactively anticipate and respond to trends, patterns and shocks.

But traditional forecasting systems simply weren’t designed to adapt to the complexity and volatility in today’s environment. They tend to be rules-based, operating in a linear and sequential way, and because they don’t account for external factors, they fall short in determining variability. These systems will never be as responsive as demand planners need them to be because of their inability to learn and account for inherent human bias.

So, where does that leave us? Isn’t technology supposed to make demand planners’ lives easier, not harder? My hope is these questions will help guide you and your organization to a more agile and scalable forecasting methodology:

  • Are you focused on generating statistical forecasts or providing AI-based operational forecast guidance to demand planners?
  • How deep and context aware is your forecasting methodology?
  • How anticipatory is your forecasting?
  • Does your forecasting approach incorporate human judgement while removing forecast bias?
  • How agile and scalable is your forecasting methodology?

 

Are you focused on generating statistical forecasts or providing AI-based operational forecast guidance to demand planners

Traditional statistical forecasting has served demand planners well until now. The usual process of demand planning starts with generating a statistical forecast mostly based on traditional time series methods. Significant effort is put into to generate these forecasts – cleaning history for impact of different factors, data quality assurance, using different models to generate the forecast and picking the best model to suit the requirements. Depending on the company, these tasks are either done by the demand planner or an in-house or outsourced statistical forecasting team.

Once the statistical forecasts are generated, then the real work of enriching these forecasts is undertaken to get the forecast to improve as much as possible.

Manual enrichment is the most value-additive step of the forecasting process. This takes the form of a) taking inputs from sales & marketing teams, b) enriching the forecasts manually based on the inputs and judgement, c) agreeing on consensus numbers at a category/brand/national levels, d) Disaggregating and reconciling the numbers to the lowest item-week-week/month.

The output of the above process is the final “operational” or “enriched forecast” that drives the upstream and downstream execution.

Many forecasting vendors and engines pride them-selves on the statistical forecasts.  However, the real forecast to improve upon is the operational or enriched forecast. That is where the value lies in both business RoI, as well as operational effectiveness and efficiency. The real test is whether the forecasting vendor’s approach can reliably beat the company’s current operational forecast accuracies.

 

How deep and context aware is your forecasting methodology?

In today’s world, Deep Learning has broken the frontiers of forecasting accuracy compared to traditional time series and even machine learning models. This is clearly seen in the latest Kaggle Competitions and real-world implementations in select advanced companies. 

Today utilizing deep learning models is not only advantageous but also necessary to learn the complex interactions, patterns and non-linear relationships in demand patterns and their movements.

Deep learning algorithms are also able to better handle some of the complex phenomena in the CPG world such as moving holidays, complex seasonality, new and innovation product introductions, seasonal products, etc.

Further Deep learning models can capture multi-level trends and patterns and improve forecasts across levels (how are category growth trends impacting operational SKU-location level forecasts, etc.)

However, getting deep learning methods right requires deep specialization in the following areas:

  1. Network architectures
  2. Hyper tuning of network architecture parameters such as Learning Rates, Number of Layers, Optimizer, Scheduler, Batch Size, Momentum, etc.
  3. Optimizing computational strategies
  4. Robust regularization and validation

Although the libraries for using deep learning algorithms are available in the open source world, what differs is how well the implementation is able to tune the deep learning architectures to make them useful to get the 10X advantage.

There is a big difference between mediocre implementation of these algorithms vs. highly purpose-built and scientifically well engineered implementations.

 

How anticipatory is your forecasting?

Many forecasting solutions are able to bring in internal and external factors but fail when future information for these factors are not available. (For example, retailer promotions, consumer promotions, trade promotions, etc.) In many cases the information from the trade, sales and marketing teams does not reach demand planners in time to be used appropriately.

As a result, for such factors historical data might be available for past history but not for the future.  This extends to other external factors as well, depending on the situation, such as weather, raw material prices, etc.

In such cases how does the forecasting methodology adapt? Does it exclude such factors or is it able to build in anticipatory methods to account for the impact of such factors?

This becomes a crucial difference for forecasting success.

 

Does your forecasting approach incorporate human judgement while removing forecast bias?

Human bias in forecasting is a well-known problem. Planning overreaction and over-optimism bias shows up in either consistent over-forecasting or under-forecasting in many CPG categories (for example, beauty products, new product introductions, etc.) Sources for bias can come from various inputs such as individual demand planner bias, S&OP inputs, etc.

At the same time human intuition, judgement and access to qualitative information cannot be ignored. Any new age forecasting methodology should be able to tease out the human bias noise from the human judgement signal and bring the best of Human Intelligence and Artificial Intelligence together.

 

How agile and scalable is the forecasting methodology?

In today’s highly volatile environment the test for a forecasting technology is whether it can adapt to changing forecasting requirements and support the increased speed of planning cycles. The current COVID-19 scenario is one case in time. Many organizations are shifting from one a month to once a week and twice a week planning and forecasting cycles and doing it for shorter time step horizons (from the next 3 months to the next 4 weeks to the next week, etc.).

Your forecasting technology should not only be able to support all of these requirements individually but must be able to support different forecasts for different horizons and frequencies in a reconciled manner.

Further the technology should be able to adapt to both the scale of the product-location-time bucket combinations and the required response times – overnight, over the weekend, etc.

Architecting and designing your forecasting technology and pipelines for scale, speed and agility is crucial.

 

Capture Revenue Growth Potential With Dynamic Demand.ai

Welcome to agile demand planning. Our engine blends intelligent cognitive automation and collaborative workflows to spot human bias, create transparency, and uncover previously hidden probabilities that will impact future inventory, promotions, and sales.

Privacy   .  ©2020 samya.ai