Introducing MMF Agent for demand planning teams
by Ryuta Yoshimatsu , Puneet Jain, Lourdes Angélica Martinez Medina, Lucas Bruand and Dael Williamson
*Enterprise demand forecasting has grown too complex for legacy tools—millions of time series, SKU proliferation, and tight planning cycles have outpaced both the technology and the talent available to run it
*MMF Agent is a guided AI workflow built on Genie Code that makes Databricks' multi-model forecasting framework accessible without requiring deep data science expertise
*Teams using MMF Agent compress days of setup into hours, produce cleaner training data, and unlock multi-model accuracy improvements that were previously limited to organizations with specialist forecasting talent
Demand forecasting has always been at the center of retail and CPG planning. It shapes inventory decisions, informs production schedules, drives trade promotion investment, and sets the conditions for every S&OP conversation that follows. When the forecast is wrong, costs accumulate quickly, leading to stockouts, excess inventory, margin erosion, and downstream disruption that ripples through the supply chain and commercial teams alike.
What has changed in recent years is not the importance of the forecast. It is the degree of difficulty.
A decade ago, a demand planner working with a few thousand SKUs across a handful of channels could manage forecast quality with a combination of statistical models, spreadsheets, and hard-won institutional knowledge. That world no longer exists for most Retail and CPG organizations. SKU proliferation, the explosive growth of e-commerce channels, regional fragmentation, and the rise of short-lifecycle promotional SKUs have created forecasting environments that most legacy tools were never built to handle.
Where a planner once managed hundreds of time series, today's enterprise forecasting problems routinely involve hundreds of thousands, sometimes far more. Each time series has its own seasonality profile, its own signal-to-noise characteristics, and its own sensitivity to external variables like promotions, weather, and macroeconomic conditions. The statistical techniques that served well at smaller scales simply do not generalize reliably at this volume and variety. Accuracy degrades. Exception management becomes unsustainable. The forecast loses its authority as a planning input.
The answer that most sophisticated forecasting teams have converged on is a multi-model approach: rather than selecting a single technique and applying it uniformly, you evaluate a range of models against your actual data and let the results determine which performs best for each time series. In practice, this produces noticeably better accuracy, but it also creates a new challenge.
Running a rigorous multi-model forecasting evaluation at enterprise scale is not a task that can be handed to a business analyst or a newly hired data scientist. It requires deep familiarity with statistical forecasting methods, modern machine learning and deep learning approaches, and increasingly, the class of foundation models built on transformer architectures that have emerged in the last few years as a promising tool for time series prediction. It also requires the ability to configure and operate a distributed computing infrastructure at the scale needed to process millions of time series within a planning cycle.
This expertise is scarce. Demand planning functions compete with every other part of the enterprise for data science talent, and the specific combination of forecasting domain knowledge and distributed systems fluency this work demands is genuinely rare. Teams that have it are productive. Teams that lack it find themselves either locked into legacy approaches that underperform against modern alternatives or dependent on a single expert whose departure creates real organizational risk.
Even for organizations with the talent, setting up a proper forecasting experiment (configuring compute resources, preparing and cleaning data, selecting evaluation parameters, running backtests, interpreting results) can stretch into days or weeks before a single model comparison is completed. In a planning environment where the cadence is measured in weeks, that cycle time is often simply too slow.
In 2024, Databricks released Many Model Forecasting (MMF), an open source framework built for large-scale, multi-model time series forecasting. MMF integrates more than 35 forecasting models drawn from leading open source libraries, including statistical approaches from statsforecast and sktime, deep learning models from neuralforecast, and foundation time series models from Chronos and TimesFM. Rather than committing to a single technique, MMF allows teams to evaluate all techniques simultaneously against their own data, with the best-performing model automatically selected for each time series.
The framework runs natively on Databricks, using distributed compute to process the volume that enterprise Retail and CPG forecasting problems demand. Dozens of companies now run production pipelines on MMF for planning decisions that directly affect revenue and inventory investment. Accuracy improvements and reductions in manual forecasting effort have been consistent findings across those deployments.
But MMF remained a tool for experts. The barrier was never the framework itself. It was the depth of knowledge required to set it up correctly, to make sound decisions about compute configuration, data preparation, and evaluation design, and to interpret the results in a way that could actually inform planning decisions. MMF made expert-level forecasting faster. It did not yet make it accessible.
MMF Agent addresses that gap. Built on Genie Code, Databricks' AI coding assistant, MMF Agent wraps the MMF framework in an interactive, guided workflow that takes teams through the entire forecasting pipeline from raw data to deployed forecast, without requiring deep technical expertise to operate.
The agent works through five stages. It begins by examining the input data, identifying quality issues, handling missing values and anomalies, and ensuring that everything is structured correctly for the forecasting engine. It then profiles and classifies the time series in the dataset, separating forecastable series from those with insufficient signal. This step is easy to skip when running MMF manually, but it consistently improves both accuracy and computational efficiency by directing resources where they will have the most impact. From there, the agent configures the appropriate compute infrastructure for the models being evaluated, executes the forecasting jobs, and performs post-processing and model selection, presenting results in terms that connect to the business outcomes the planning team actually cares about. The short walkthrough below shows a demand planner moving through each of these stages with the MMF Agent in Genie Code.
What makes this different from simply automating a workflow is that the agent is interactive. It draws on Genie Code's integration with Unity Catalog to understand the organization's full data environment, enabling it to make informed recommendations on which datasets to use, how to enrich training data with relevant external variables, and how to interpret forecast accuracy metrics in business terms. A planning leader who understands their business but is not a data scientist can engage with MMF Agent in the language of demand planning (promotions, seasonality, channel mix, planning horizons) and receive guidance grounded in both forecasting best practice and the specifics of their data.
The most immediate benefit is speed. Setup and experimentation work that previously required days of skilled data science effort can be completed in hours. That compression matters in a planning environment because it means teams can run more experiments, test more model configurations, and respond more quickly when market conditions shift. This is critical when historical patterns stop being reliable guides.
Forecast accuracy also tends to improve. The data preparation and series classification steps that MMF Agent guides teams through produce cleaner training data and better-targeted model selection than manual approaches typically achieve.
Perhaps the most consequential change, though, is reach. Demand planning teams that lack dedicated forecasting data scientists can now operate with a level of methodological rigor that was previously out of reach. The expertise barrier that kept multi-model forecasting confined to organizations with specialist talent is no longer what it was, opening the approach to a much wider range of mid-market Retail and CPG organizations.
For teams that do have strong technical depth, MMF Agent also makes the framework easier to customize. MMF has always been open source, but in practice, only a handful of teams have had the expertise to modify it. When the agent has access to both the source code and the guiding skills, it can walk engineers through changes in plain language: adding a new model class, adjusting the backtesting logic, and integrating a business-specific accuracy metric. Modifications that once required deep framework knowledge become approachable for a much broader range of engineers.
MMF and MMF Agent are available today. The MMF Agent skills are available in the Many Model Forecasting GitHub repository, along with documentation and example notebooks covering the full workflow. Install the skills into Genie Code or a local agent environment, and your assistant will take it from there.
For demand planning leaders who want better forecasts with the team and tools they already have, this is worth an hour of your time to explore.
Subscribe to our blog and get the latest posts delivered to your inbox.