A No-Code ML Forecasting Platform for Retail and CPG Companies

Summary

• ML forecasting simplified: Low-code/no-code platforms enable retail and CPG data scientists to build, train, and deploy ML forecasting models in hours instead of months.
• Business ROI aligned: Accurate demand forecasting directly impacts inventory planning, pricing decisions, product placement, and profitability across retail operations.
• Enterprise scalability: Platform-based solutions accelerate time-to-value by abstracting away engineering complexity while maintaining model quality and governance.

A No-Code ML Forecasting Platform for Retail and CPG Companies

Talk by Moez Ali, AI Product Director, Zebra Technologies
Retail and CPG companies face growing pressure to better forecast demand, optimize pricing and manage inventory — yet traditional approaches take months to deploy and often require extensive engineering support. In this video, we will showcase Workcloud Modeling Studio, a low-code/no-code ML platform designed for data scientists working in retail and CPG. Learn how this tool improves forecasting accuracy and accelerates time-to-value from months to hours. We will walk through a real-world use case of demand forecasting for a retailer using Zebra's Modeling Studio. This talk will demonstrate how to build, train and deploy an ML forecasting pipeline — without reinventing the wheel.
Databricks Named a Leader in the 2025 Gartner® Magic Quadrant™ for Data Science and Machine Learning Platforms: https://www.databricks.com/blog/databricks-named-leader-2025-gartner-magic-quadrant-data-science-and-machine-learning
Build and deploy quality AI agent systems: https://www.databricks.com/product/artificial-intelligence
See all the product announcements from Data + AI Summit: https://www.databricks.com/events/dataaisummit-2025-announcements

FAQs

Why do retailers need ML forecasting?

Accurate forecasting is foundational to inventory optimization, pricing strategy, and profitability. According to KPMG, 60% of CPG executives cite accurate forecasting as their top challenge to achieving a profitable and resilient supply chain.

How does no-code forecasting reduce deployment time?

No-code platforms abstract complex ML workflows, allowing data scientists to focus on data preparation and model evaluation rather than infrastructure. This reduces deployment from months to hours, enabling rapid iteration and faster ROI.

What is an enterprise machine learning platform?

An enterprise machine learning platform provides data scientists and analysts with pre-built algorithms, data preprocessing, and deployment tools to build production models without extensive engineering overhead. This video demonstrates Workcloud Modeling Studio, which sits on top of the Databricks Data and AI platform and provides a no-code interface for building forecasting pipelines. The platform abstracts infrastructure complexity while maintaining access to sophisticated ML capabilities for organizations across retail and CPG industries.

What is the best enterprise machine learning platform?

The best enterprise ML platform combines ease of use with production scalability, which this video illustrates through Workcloud Modeling Studio built on Databricks infrastructure. The platform enables data scientists to build, train, and deploy ML forecasting models in hours instead of months by providing pre-built components and automating infrastructure management. By keeping data within your Databricks account while offering a single-tenant UI, it balances accessibility with security and governance.

Full transcript

All right, good afternoon everyone. Thank you for joining me today. I'm a product manager at Zebra and I'm really excited here to be sharing the work that uh that my team is doing for the last couple of years. So, Zebra has been enabling the front line of uh retail and CPG companies for over uh 50 years now. Um we have worked with more than 80% of Fortune 500 companies. We have worked with almost every large retailer uh in the world.
You probably know Zebra from its hardware devices like scanner, printer, handheld. But since uh 2019 uh Zebra has started a portfolio started to acquire a portfolio of software companies uh so that we can increase the value that we deliver to our customers and I came to Zebra as as part of that acquisition in 2021 through nuai and what used to do is build forecasting solutions for retail and CPG companies. So let's start with the why, right?
This is a survey from KPMG uh where 60% of CPG executives feel that accurate forecasting is the number one challenge preventing them to achieve a profitable and resilient supply chain. And if you look at the other challenges on this slide as well, they are all mostly related to accurate forecasting and kind of makes sense that uh if your forecasting is is bad, all the downstream decision making uh is going to get affected.
Now what do if we are able to achieve accurate forecast what do retailers and CPG do with these forecast? Right? It turns out that uh it turns out that uh forecasting is used across the board for decision- making whether it's inventory planning, manufacturer planning, whether it's marketing and promotion, pricing decisions, uh product placement or even new product launches. Forecasting is the underpinning ingredient uh to you know convert some of these use cases into six figure, seven figure ROI for for your companies. forecasting has fundamentally changed over last decade.
We have gone from really high level static batch weekly forecast uh powered by limited computing to you know this infinite compute cloud scalable skew store daily level forecast right and um as we move from you know forecasting in 2025 has become uh a hyper techchnical exercise and we see a problem in that because the more technical forecasting becomes the further it moves moves away from the business, right? And even with companies with very resourceful and really large great data science teams, uh building accurate forecasting model at a scale, uh and be able to maintain the accuracy over a long period of time is is far from easy, right?
And it's not because that algorithms and models that we have today doesn't work. It's because the disproportionate amount of time teams are spent in setting up infrastructure, tuning infrastructure, doing deployment, MLOps, everything except for thinking about how can we maximize the business impact from these accurate forecast and this is the problem that we are sort of on we this is the problem that we are trying to solve uh with our platform called workcloud modeling studio.
Workcloud Modeling Studio is a no code, low code machine learning platform targeted for data science teams in retail and CPG companies to be able to train, test and deploy highly scalable machine learning forecasting pipelines. Now looking at this screen, the tool looks pretty intuitive. Almost all the workflow tools work like this. You visually build a pipeline except there one thing you would notice on the left hand side that you see those small boxes those are steps or operations that you can perform uh in modeling studio right so the idea is very simple u the tool comes the application comes with lot of algorithms pre-processing steps feature engineering uh that data science teams can use to visually build this without writing any single line of This is a highlevel blueprint uh of the platform.
In the middle you see uh first thing you'd see the platform sits on top of your data bricks. So if you are a datab bricks uh user is spinning up modeling studio is going to be very easy. The data and the infrastructure uh is provided by data bricks right. So your data never comes out from your infrastructure. Uh and whatever you do in modeling studio is is single tenant zebra hosted UI but underneath it is going to use your existing data bricks and data lake infrastructure.
What you are seeing on right are different use cases of forecasting. Once you generate a forecast you'd be able to sort of use them for allocation replenishment or all these other use cases that you have. All right, with that I am going to change my screen to show you the platform. This is the landing page of the application. You can see a search experience on the top. You can see documentation links on the right, some recent uh in the middle.
Uh you have leftand navigation on the left hand side. For the first time when you start datab bricks uh you would have to go and create a developer token in your datab bricks account and pass that developer token here. That's how we are able to sort of create the compute and everything in data bricks on your behalf. And this is just a one-time thing. But once you have done that you can the first thing normally people would do is create a compute resource here.
This experience is pretty similar to data bricks. You can simply click on a create compute button, give a name, set a resource and uh you can click on create. As you can see, I have uh few compute already and two of them are running, right? And if you go to your databicks account, you would you would be able to see logs uh in real time for these compute because these are actually running in your data bricks account, right?
Once you have compute set up, uh what you do is the work in modeling studio is organized in projects. Project is the highest level in hierarchy. All the data, experiments, pipelines, artifacts belong to a project and I have couple of projects here, right? Um and all the work that you do, right? Can be either um only limited to you or you can share with AC across other teams in your organization. So the tool is built to support uh the collaboration uh and partnership across teams as well.
So what I'm going to do is I have a project here called data bricks demo right and it has six steps. It's a I've built this three days ago. Uh this is a anonymized data set but it's a forecasting use case where I have data set of actual skew store level uh with lot of promotion signals calendar data and all this different sort of data. The project is organized. It's a typical way to organize a project but it's up to user. uh but there are six steps in this process starting from pre-processing the second step is feature engineering then we are training a very simple light GBM model uh then we are predicting it and then finally we are preparing the uh final output right so if I open the first experiment so we call this screen canvas right uh you have seen this before on the left hand side you have all these different uh steps we call them transformers they are the original implementation on spark right so what is really unique here that this is actually running off your compute in data bricks and all the implementation whether it's join table or filter table whether it's a light GBM model whether it's a uh promotion model whether it's a time series decomposition or ARMA everything is natively implemented in spark so that it's very costefficient for you and it's scalable right So this pipeline is very generic uh just joining couple of data sources.
I'll show you a few of them. Example this one, right? This is my calendar table. You can see I have all these different events here. 75 different columns, 1900 rows. I have a uh just a hierarchy item department category hierarchy here. Uh and I believe this is my f data. You can see this is 59 million records. It took me about two and a half hour to run the entire thing that produced the forecast. Uh very good forecast indeed.
So this step uh is very generic uh pre-processing just filling missing values uh joining tables and basically preparing the data set uh for the next step. If I bring up step two here, this is a little bit more data sciency. So what this is doing uh the two main things sorry I I opened step one again. So the two main things that uh this step is doing is you know it is very common that you have lot of zeros in your data in your sales data and that may not necessarily mean no demand.
It may mean out of a stock right? So this is a pretty common technique where you try to detect what are the days where you were out of the stock because your goal is to predict the demand not the sales right. So we have proprietary implementation of these algorithms here that can do that job for you and create a flag in your data set so that you can then imputee them uh with some realistic number instead of zero. The other thing that is also very sort of common in retail is you don't just have units sold but you also have lot of promotional data, event data like was the was the unit on sale, some kind of promotion on that day, what was the weather like, right?
Depending on what you are forecasting, there could be lot of exogenous variables, right? And when you are using some sort of regressor model to forecast what you have to do is you have to take the effect of all those exogenous variables out predict the base demand and then join them back. Right? So this is what this step is uh mostly doing. It is basically removing the impact of all the other events that we have in our data set so that we can forecast the base demand with light GBM model and then at the last step we are going to add it back.
All right, the third step um after I have removed the effect of promotion, remove the effect of other exogenous variables. This step is simply just throwing that process data set into a light GBM model uh which gets sort of used in step four uh that produces uh the forecast. Right? So this is what we were after the final output right and you can see there's depromotionalized sales unit decomposed and the step five and six just sort of adding it back connecting it together right the really good thing about this as you are doing these experimentations in modeling studio you can all the output that you produce here is basically a artifact that you can see here and This artifact is essentially u uh a parquet file uh sitting on your datab bricks.
Right? So basically what you can do for example this is what I did. I loaded in the final output in my data bricks notebook and just visualize here. So this is the forecast that I produced at a aggregate level. The orange line is the forecast right and this is more at a item level. So anything that you produce in uh in modeling studio you can very easily switch back and forth between data bricks and modeling studio. Right.
The other thing I would say that once you sort of do your experimentation here do your model selection here. Uh it is very easy to simply deploy this pipeline at which time you know you set a schedule and this pipeline would basically run as a datab bricks job in your datab bricks account at a status schedule. If you want you can monitor them in modeling studio UI. Uh but these jobs would also be there in datab bricks account.
If you feel comfortable you can also uh monitor the jobs from there as well. All right, So if you have a forecasting uh use case in your company, if you are a data bricks customer, right, this is going to sort of we think this is going to maximize or increase the value that you are getting from uh from your infrastructure, right? And if you're interested in trying, we are giving uh free trials. You can sign up a quick form.
It would take you 10 second and somebody from our team would reach out and we are happy to set a more detailed data sciency demo and we are happy to uh give you a trial to try it out. Thank you so much everyone.

Learn more about the Databricks Data and AI platform.