Updated on June 10, 2026
Most supply chain teams know they need better forecasting. What’s less understood is that better forecasting doesn’t come from finding the one best algorithm — it comes from using the right algorithm for each situation.
That distinction matters more than most software vendors will admit. And it shapes the entire way we’ve built forecasting into New Horizon’s demand planning software.
Facebook Prophet: A Genuine Breakthrough
When Facebook released Prophet as open-source software in 2017, it was a meaningful shift in the forecasting landscape. Facebook had originally built it for internal use — things like forecasting demand for data center computing resources and advertising inventory. The underlying engineering that powered those use cases turned out to be exceptionally well suited to supply chain forecasting, too.
Prophet handles large, non-linear data trends well. It manages trend shifts, outliers, and missing data automatically. It works effectively without a team of data scientists configuring it. Those aren’t small things when you’re running demand planning at scale.
Since its release, Prophet has been downloaded nearly 30 million times and is widely considered the most popular open-source forecasting software in the world. It’s also one of the algorithms embedded in Amazon Forecast — which tells you something about the industry’s confidence in it.
New Horizon has used Prophet as part of our forecasting algorithm library for years, and we’ve made industry-specific enhancements to further improve its accuracy for our customers.
What Our Testing Actually Showed
We ran a large-scale internal test comparing Prophet against traditional algorithms — ARIMA, Holt-Winters, and Exponential Smoothing — across 266,549 items.
The result: Prophet was the top-performing method more than 62% of the time.
That’s a strong result. But the number worth paying attention to is the other 38%. In nearly four out of ten cases, a traditional algorithm outperformed Prophet. That’s not a knock on Prophet — it’s a realistic picture of how forecasting actually works in practice.
No single algorithm wins every time. Demand patterns vary too much across products, industries, and data histories for any one approach to dominate universally. For a deeper look at how AI and traditional methods compare, see our AI for Demand Forecasting: A Primer on Smarter Predictions.
Why the Algorithm Has to Fit the Situation
Here’s the practical reality we see across our customer base.
Prophet performs best when there’s a lot of data to work with — specifically, situations where customers have years of daily or even intraday sales history. Traditional retail and online retail environments are ideal use cases. With high-frequency data, Prophet’s ability to identify complex seasonal patterns and trend shifts is a genuine advantage.
But many companies don’t have that kind of data depth for every product. In consumer goods and foodservice, for example, it’s common to work with two years of monthly data — roughly 24 data points per item-location combination. In those situations, ARIMA and exponential smoothing typically produce better forecasts. For weekly data covering two years (about 104 data points), the right algorithm varies case by case.
Highly seasonal demand calls for different models than steady demand. Intermittent demand calls for different models than continuous demand. The amount of historical data shapes the choice as much as the demand pattern itself.
This is why we use more than 20 different forecasting models in our Demand Planning application — some traditional, some newer AI-driven models, and some cutting-edge approaches that use artificial neural networks. The goal is to have the right tool available for every situation, and then use AI to automatically select the best one.
AI-Driven Model Selection: The Real Differentiator
Choosing among 20+ forecasting models manually — for every product-location combination in a large catalog — isn’t practical. That’s where AI-driven model selection comes in.
Our system automatically evaluates which forecasting model will produce the most accurate result for each specific SKU and location combination. Every product gets the model that fits its demand pattern and data profile, not the model that performs best on average across the catalog.
This is what separates a truly AI-driven approach from one that simply uses an AI algorithm. Using Prophet everywhere is an improvement over using ARIMA everywhere. But automatically selecting the right model for each situation — and continuously updating those selections as demand patterns evolve — is the approach that actually maximizes forecast accuracy across a diverse product portfolio. It’s also why purpose-built AI-embedded supply chain planning software consistently outperforms custom-built alternatives that rely on a single algorithm or approach.
Transparency and Planner Control
Automation is only useful if planners trust it. That’s why we’ve built explainability into how our forecasting works, not just accuracy.
Planners can see which forecasting model the system selected for any given product, and why. They can see the components of a forecast — the baseline trend, the seasonal pattern, the impact of causal factors like promotions or weather — rather than just a single output number. And they can override the AI-selected model and choose their own if they have information the system doesn’t, such as a market shift or a product change that hasn’t shown up in the data yet.
The system advises. The planner decides. That balance between automation and human judgment is essential for planners to actually use and trust what the software produces. We go deeper on explainability and the other key things to know about AI in planning in 5 Things You Need to Know About AI and Supply Chain Planning.
What’s Next: NeuralProphet
The forecasting technology landscape continues to move quickly. NeuralProphet — released by Facebook in collaboration with researchers at Stanford University and Monash University — builds on Prophet by adding deep learning capabilities through AR-Net, a neural network algorithm developed by the Facebook AI Research group.
It’s particularly well-suited to causal factor forecasting (where variables like weather or promotions have a significant impact on demand) and to situations with sparse or intermittent demand data. Its developers report forecast accuracy improvements of 55 to 92% for short- to medium-term forecasts.
New Horizon is evaluating NeuralProphet for inclusion in our Demand Planning application. As with every algorithm we add, the question isn’t whether it’s impressive in isolation — it’s whether it belongs in the portfolio and which situations it serves best.
The Right Algorithm for Every Product, Every Time
Prophet is a powerful forecasting algorithm, and we’re glad to have it in our toolkit. But the real advantage for supply chain planners isn’t any single algorithm — it’s having access to a broad library of models, with AI selecting the right one automatically based on the actual characteristics of each product and location.
That’s how you get the accuracy gains at scale that matter for real supply chain performance. For a closer look at how machine learning scales across complex demand planning environments, see our post on scaling demand planning with machine learning.
To learn more about how AI and machine learning work in demand planning — including where AI adds the most value and where traditional methods still hold their own — read our AI for Demand Forecasting primer. Better yet, contact us to talk through how New Horizon’s approach could work for your planning environment.