Forecasting and large language models are currently top of mind for many data‑science teams. In areas such as demand and material planning, production planning, and inventory optimization, accurate forecasts deliver measurable gains in efficiency and reliability along the supply chain. In recent years, generative AI (GenAI) has become a key technology that unlocks new possibilities. Using demand forecasting as an example, we show how GenAI can not only improve forecast quality, but also make the topic more accessible and understandable to a wider audience.
The status quo in demand forecasting
Data analysis, pattern recognition, and the value of reliable forecasts
Forecasting is an essential practice across the supply chain and especially impactful in demand planning. Accurate forecasts of material needs, product sales, and revenue lead to more precise planning and help detect potential trend reversals early. In short: better service levels and lower inventories, because production aligns with actual demand.
Methodologically, demand forecasting typically analyzes historical consumption for patterns such as seasonality and trends, which are then projected forward. This analysis and forecasting relies on established methods from statistics and machine learning – from classical time‑series techniques like ARIMA to more complex ML models. Applied thoughtfully, they perform well.
External influences on demand: the challenge of identifying and integrating relevant drivers
In addition to historical consumption patterns, demand is often influenced by external factors such as economic indicators, weather, holidays, or marketing activity. These are frequent drivers of turning points or sudden peaks and dips. It’s natural to want to include them in forecasting; the challenge is first to identify the relevant drivers and then to integrate them properly into the models.
Summing up, robust demand forecasting needs, among other things:
- models that produce forecasts from data,
- suitable drivers to feed into those models, and
- good explanations to create transparency and trust.
GenAI can be a game changer on all three fronts. Let’s look at how.
How to create forecasts with generative AI
Global time‑series models
Generative AI – and LLMs in particular – are best known from tools like ChatGPT, Copilot, Gemini, or Claude. Those systems were trained on large text corpora and learned to continue text. In forecasting, the “sequence” is a time series: numbers ordered in time. If LLMs can learn from text sequences, why not from numeric sequences too?
In fact, transformer‑based global time‑series models exist and have been trained on vast collections of series. One example of such a foundation model is the open‑source model “MoraiAI”. Their promise: produce forecasts in seconds, without having to train a bespoke model with individual coefficients for each series. They can also be attractive where history is short and robust statistical models are hard to fit.
Views on these models are still split. Some experts doubt transformers can handle time series with the same precision as language; others are confident transformer‑based methods will soon complement or even replace classical approaches like ARIMA or exponential smoothing. Our own initial tests, for example on demand forecasting for sales volumes, have shown encouraging results – though our first conclusion is that the approach still needs tuning.
A pragmatic compromise could be to train transformer models specifically for a use case like demand forecasting rather than one universal model for all series. That way, we retain the benefits of the architecture while keeping accuracy high for the problem at hand.
Chatbot‑assisted forecasting
It gets especially interesting when we combine LLMs’ language skills with forecasting models: imagine a forecasting assistant you can chat with and that returns the requested forecasts, analyses, visualizations, and explanations right away. No special software to learn, no data‑science background required – yet useful results:
“I’m giving you a monthly sales time series. Please forecast the next 12 months.”
Behind the scenes, the assistant applies suitable techniques, compares candidates, and returns the outputs – numerically and as charts.
An easier, more intuitive way to access forecasting is hard to imagine.
How to find suitable drivers with generative AI
We’ve looked at how GenAI can help create forecasts. Now let’s explore how to make them better by bringing in drivers that truly explain future demand – not just the target series’ history.
Evaluate drivers: separate causality from spurious correlation
Consider the example of rainy days in San Francisco versus the number of printing‑press operators in Rhode Island.
If you only look at the numbers, you’ll find a correlation. But there’s no logical or causal link between the two. With LLMs, we can finally bring context into the evaluation. They can draw on general knowledge, industry knowledge, and even internal documentation to answer: is an observed correlation spurious, or is there a plausible causal relationship?
For the example above, we asked Meta’s Llama to reflect and critique its own answer. Its assessment: “It appears that the correlation is most likely a spurious correlation.” Correct.
Automating this kind of screening across many candidate drivers for a given time series is feasible – and even if the answer isn’t always black‑and‑white, you’ll move much faster toward a defensible short list.
Preselect drivers: reduce many unspecific drivers to a few relevant ones
A similar technique helps when there are far too many potential drivers. Clients sometimes hand us databases with tens of thousands of candidates. Running them all through models is often infeasible. Instead, we can use LLMs to reason over textual metadata (name, definition, unit, domain, source, region) and identify a smaller set that actually fits the problem. Only that subset then goes into quantitative evaluation – saving time and compute power.
Construct drivers: turn text into new explanatory variables
Some of the most impactful influences are one‑off or rare events that live only in text: floods affecting agriculture, a factory fire disrupting a supply chain, a regulation change. With grounding and web‑scrolling, LLMs can retrieve such unstructured information and transform it into structured, time‑aligned signals that can be analyzed alongside the target.
Even when you can’t fully convert text into numeric signals that lift accuracy, these insights are still invaluable for interpretability and explanation.
Example: “Why did demand drop so sharply in 2008? Were there events in 2008 tied to our industry that could explain it?” GenAI models can help explain such anomalies using textual sources.
How to communicate results with generative AI: personalized reporting
GenAI also shines when it comes to explaining and communicating results. Many users find forecasting outputs complex or opaque. Offering a way to ask questions in their own words – e.g., through a tailored chatbot – can make a decisive difference for acceptance and trust.
Examples:
“What explains the higher forecast for May 2025 compared with last year?”
“Did the model include a trend?”
GenAI can also tailor outputs to audiences. Management might prefer quarterly summaries; demand planners prefer weekly or daily detail. A conversational interface can generate on‑demand reports and charts in the exact format required.
Conclusion: GenAI is a game changer for forecasting
GenAI has the potential to fundamentally improve forecasting. Foundation models provide new forecasting methods that are likely to become more accurate, faster, and more robust. In addition, GenAI advances explainability and brings unstructured textual context into analytical workflows. That combination unlocks practical innovation – especially in demand planning.
It’s also a powerful aid for data scientists working across many industries: GenAI helps us get up to speed faster in new domains and dig deeper into context.
We’re confident that, thanks to its usability and accessibility advantages, GenAI will become integral to forecasting applications.