A global pandemic, a fundamental shift in the demand for goods and services worldwide, and the recent blockage of a major international trade route have all highlighted the need to build and maintain resilient supply chains.

At the foundation of resilient supply chains lie accurate and reliable forecasts. The majority of traditional software systems rely heavily on domain-specific rules; the disruptions due to the pandemic have broken the existing logic and rendered such systems unusable.

In this article, we discuss the problems with traditional supply chain forecasting and how AI can be used to help. We show how the combination of AI and cloud computing makes forecasting significantly faster, more adaptable to market changes, and more precise. Not only that, we demonstrate that when it comes to supply chain management for goods like medical supplies or food, AI is not a nice-to-have but a must-have.

The Problem – More Disruption More Often

The pandemic created an unprecedented shock to our supply chains and economic environment. We are witnessing a historic disruption of both the supply and demand sides of the economy, as well as policies enacted as the world attempts to find a new normal.

Early on, many thought demand would drop as people began to prepare for dire financial conditions. Yet, when manufacturing hubs had to shut down, people began to hear of supply shortages, which led to a dramatic spike in demand for countless goods as countries prepared for lockdowns.

These sharp spikes and drops in demands for goods and services have lasting effects with respect to production, trade, and prices. Today we are witnessing the tailwind of this volatility as markets look to return to their normal equilibriums. Not to mention the less obvious but continuous impacts of climate change.

Disruption is rife and shows no signs of stopping.

Industry Examples

Let’s go through some examples by industry to give you a flavor of typical supply chain forecasting examples.

Procurement

Supplier Management  – how do you ensure your business receives maximum value from your suppliers? Can you cope if they have shortages? Can you predict their shortages and order more stock ahead of time to ensure sustainable supply?

Demand Sensing  – how can you react to daily, hourly, or even real-time changes in the supply chain?

Cost Tracking  – with all these disruptions, will you have to pay top-dollar to get your goods delivered through expensive methods? Or can you predict these issues ahead of time and keep costs down?

Manufacturing

Predictive Maintenance  – can you accurately track your assets’ health in real-time and schedule maintenance around your production schedule? Or is downtime happening unpredictably, resulting in lost working hours and falling behind schedule?

Process Optimization  – are you wasting time at your factories with inefficient processes? Can you confidently implement new procedures knowing they will lead to efficiency gains? Or are you left in the dark about your process optimization?

Distribution

Inventory Management  – how do you ensure your products never go out of stock? What is the optimal trade-off between spoiled inventory and sales?

Route Optimization  – which routes should your transportation staff take? Can you integrate real-time traffic updates?

All of the examples listed above have been solved using traditional methods for years. And they are all prime candidates for AI disruption; they are closed problems with clear objectives and plenty of data to work with.

Note that supply chain forecasting problems can involve more than just structured tabular data. Some involve networks (route optimization); others can include text, image, and video data as well.

Up until now, many businesses have been using traditional supply chain forecasting approaches and have had satisfactory results. But covid changed everything, and the issues with these techniques have been starkly brought into the light. Let’s look at some of them now.

Challenges with Traditional Approaches

The traditional supply chain forecasting approaches (looking at you excel spreadsheets!) have experienced a wave of challenges over the last few years.

Broken Assumptions 
The biggest and most important of these is that the underlying assumptions underpinning the models have changed. In this post-covid era, many variables are constantly changing, and we have limited historical data to help inform our decisions. This is a massive problem. We need data to build forecasts, and this restricted amount is pushing every model to its limits. Simpler models are less good at extracting data than AI, and so these models struggle to produce accurate forecasts.

Slow Model Development 
Building traditional models involves much manual work. They require time and expertise to create, and this results in sluggish model development. In this swiftly changing environment, we need to build thousands of models each day and adapt to daily changes. Therefore, slow models are close to useless.

This is not just a problem for traditional methods, however. Even if you have a Data Science  team building models for you using AI, you are probably still operating too slowly. It can take weeks for a DS team to create a high-performing model. This speed is still far too leisurely for today’s climate.

Economic Cost 
There is a substantial economic cost to using poor-quality forecast models. If you have inaccurate forecasts and get hit with an unexpected event, it can devastate your supply chain and your business.

Traditional models are not cutting it anymore. AI is the future, and it is here to stay. But building, deploying, and maintaining a successful suite of AI models has its own challenges.

Challenges with AI

Choosing to use AI for your supply chain forecasts will be an improvement. But it is not without its problems. To build a world-class forecasting function, here are some things you should keep in mind.

Training Scale 
The time it takes to build and deploy one model is reduced by several orders of magnitude if you use the right AI tools. But you need to focus on building hundreds of thousands of such models. Moreover, you need to update and retrain them regularly and ensure they are performing at their best.

Model Explainability 
One risk with AI is that your solutions become black boxes spitting out predictions that no one can understand the reasoning behind. Thus, alongside the vast training scale, you must ensure all your models can be explained to anyone who wants to know.

Forecast Scale 
Different users require predictions at different levels. C-Suite executives may want country-level forecasts, but those working on the ground will want granular predictions they can immediately use. Some examples from various industries:

Sales Forecasts 

• Country → Department → Store → SKU (daily, weekly, 30-day, 90-day)

Machine Failures 

• Group of sites → Manufacturing site → Assembly line → Single Machine (hourly, daily, weekly)

Delivery Time Delays 

• Country → State → City → Department → Product (daily, weekly, monthly)

As you can see, you not only need to build different models for different datasets, you also need different levels of abstraction in your predictions and to serve them to the right stakeholders at the right time.

Deployment 
There are so many things to keep in mind for deployment that it could be an article all in itself. But some essential components to get right would be:

• Reproducing feature engineering and scoring pipelines across groups.
• Deploying your models on your platform of choice.
• Ensuring your models run successfully given the memory constraints of your platform.
• Deciding on real-time or batch scoring for your predictions. Similarly, deciding how much latency you can tolerate for your predictions.
• Monitoring model drift and implementing retraining when necessary.
• Democratizing AI throughout your business to ensure everyone can use and benefit from the model’s predictions.

This is a critical phase in the AI model building lifecycle. If you have the best models in the world but poor deployment, the models will be useless to your business. Do not scrimp on this section.

Essential Inputs

You should feed as much data as possible into your modeling pipeline and let AI pick the best features with high predictive power. We want to highlight two classes that have become prominent over the last few years that we think are essential to include.

Covid-19 Data 
The elephant in the room. Covid has been the biggest driver behind supply chain disruption for the last 1.5 years and will continue to play a huge role moving forward. You must incorporate Covid-19 data into your supply chain forecasting models. Critical components would include lockdowns, vaccination rates, rules regarding vaccinated and non-vaccinated individuals, mandatory social distancing policies, growth of new cases, and prominence of new variants.

H2O has worked hard over the last year on global covid forecast Kaggle competitions and placed in the top spots. We’ve put everything we’ve learned from these competitions into our product so that you can have access to world-class covid estimates and use these to enhance your supply chain models.

Macroeconomic Data 
We’ve seen significant changes across the global economy, and these show no signs of stopping. Key sources include unemployment levels, inflation rates (using multiple indicators), and mobility data (what percentage of people are moving to/from rural areas?).

Types of Forecasts

There are a few different types of supply chain forecasts you can make. Traditionally they have been split into two categories: short term and long term. But we also have a new one: demand sensing.

Short Term 
Short-term predictions are made for the following few weeks or months.

Some effective short-term models are Power Growth and SEIRD (Susceptible, Exposed, Infective, Recovered, and Deceased) models. H2O used the former with great success to place near to the top of several Kaggle forecasting competition leaderboards. A promising model based on the latter is the SEIRD transformer which works well with Automatic Machine Learning (AutoML ).

Long Term 
Long-term predictions are made for several months or years in the future.

It is extremely tough (if not impossible) to make accurate long-term predictions right now. There is no historical data we can draw on to determine if, for example, another wave of infections will occur and if countries are likely to implement lockdowns again.

However, all is not lost. You can use what-if simulations to test your supply chain in a range of different scenarios and see how your estimates vary. This process gives you a better understanding of the robustness of your forecasts.

Demand Sensing 
We use demand sensing to make predictions for the coming hours, days, or 1-2 weeks.

Short-term predictions may be suitable for a few weeks. But the world changes daily. To adapt to these conditions, you need to update your forecasts daily, hourly, or even in real-time. Enter demand sensing. In effect, you superimpose what happened last week onto your short-term models and alter the predictions based on the new data. This tool is crucial for retailers who need to do this daily for thousands of products in hundreds of locations.

Conclusion

We are living in unprecedented times. The world is changing daily, and this is having huge ramifications on global supply chains. Traditional models have got us this far, but they will not take us much further. It is time for AI to step in and revolutionize the way supply chain forecasting is done. You could use traditional data science teams, but it often takes weeks to build a single model, and that isn’t fast enough in a world where significant daily changes are the norm.

To test thousands of models in hours, deploy them within minutes and share them with your entire company, start a 14-day free trial  of H2O AI Cloud .