11-17-2025
When you open your food delivery app at 7 p.m. on a Friday, expecting your usual 30-minute wait time, you're witnessing the culmination of thousands of real-time forecasting calculations happening behind the scenes. Each geographical region, each time of day, each shifting pattern of consumer behavior — all feeding into algorithms that must predict demand with remarkable precision.
But what happens when the world changes overnight? COVID-19 sent shockwaves through every on-demand platform, transforming quiet suburban areas into delivery hotspots while urban centers went silent. Traditional forecasting methods, designed for stable patterns, buckled under this instability. A new approach from a Purdue University researcher and coauthors offers a solution that automatically adapts to these dramatic shifts — no human intervention required.
On-demand service platforms face what researchers call a "wicked problem." They must simultaneously forecast demand across hundreds of geographical regions, process data streaming in real-time and adapt to sudden environmental changes, all while delivering results fast enough to guide split-second operational decisions.
Traditional forecasting works well in stable environments. But when external shocks hit — whether it's a pandemic, severe weather or a major local event — these systems can fail magnificently. The algorithms continue making predictions based on outdated patterns, leading to driver shortages in high-demand areas and wasted resources in others. While industry-leading methods like recurrent neural networks (e.g. long short-term memory, or LSTM) or Facebook's Prophet algorithm are able to detect underlying parameter changes, they do so at high computational cost. Finding a way to efficiently detect structural breaks, or “change points,” therefore represents a significant source of value for on-demand platforms.
The research team, led by Yu Jeffrey Hu from Purdue’s Daniels School of Business, developed what they call FFUDS (Fast Forecasting of Unstable Data Streams). The framework does something remarkably human-like: it continuously monitors its own performance and automatically switches strategies when it detects that conditions have changed.
Here's how it works: Instead of just predicting demand, FFUDS simultaneously tracks the quality of its own predictions. When forecast errors suddenly spike — a telltale sign that the underlying environment has shifted — the system immediately begins combining two different approaches. It blends forecasts based on all historical data (for stability) with forecasts using only recent post-change data (for responsiveness).
Think of it like a GPS system that not only calculates your route but also monitors traffic conditions in real-time, automatically rerouting when it detects congestion ahead. But unlike GPS, FFUDS makes these adjustments across hundreds of parallel data streams simultaneously.
Testing the framework on data from a leading European delivery platform revealed impressive results. FFUDS consistently outperformed industry benchmarks across 294 UK delivery areas from 2019 to 2021 — a period that included the massive disruptions of the COVID-19 pandemic.
The performance gains translate directly to business value. Compared with competing methods, FFUDS generated an average of £1.1 million in annual savings, ranging from £18,000 when compared to the best-performing benchmark to £3 million when compared with the company’s existing standard — and that's just from improved forecasting accuracy. Factor in the computational efficiency gains, and the framework proved 45 to 5,500 times faster than competing methods. Given that most on-demand platforms operate via a cloud service provider, such as Amazon Web Services or Microsoft Azure, computational savings translate directly into dramatic reductions in operational costs.
For a platform operating in just one country, this could mean millions in annual savings. Scale that across multiple countries and the economic impact becomes transformational.
While the research focused on delivery platforms, the implications extend far beyond getting your dinner on time. The framework addresses a fundamental challenge facing any organization dealing with high-frequency, unstable data streams.
E-commerce platforms must predict traffic spikes during flash sales. Transportation networks need to anticipate rideshare demand during major events. Even financial markets could benefit from algorithms that automatically detect when market conditions have fundamentally changed.
The key insight is that instability isn't a bug in modern data — it’s a feature. In our rapidly changing world, the ability to automatically detect and adapt to new patterns has become just as — if not more — valuable than simply optimizing for historical ones. To illustrate the generality of their approach, the authors also tested it on publicly available data from a NYC-based bicycle-sharing platform, demonstrating its broad applicability across industries.
The research suggests we're entering an era where the most successful platforms won't just be those with the best algorithms, but those with algorithms that can automatically improve themselves. As external shocks become more frequent and severe, the ability to maintain operational excellence through uncertainty becomes a core competitive advantage.
For business leaders, the message is clear: the future belongs to systems that can think, adapt, and optimize themselves in real time. The question isn't whether your organization's forecasting methods will face unexpected disruptions — it's whether they'll be ready when they do.
Yu Jeffrey Hu is Accenture Professor of Information Technology at Purdue and a Distinguished Fellow of the INFORMS Information Systems Society. A world-renowned expert on AI, analytics, and the digital economy, he currently heads the Daniels School’s Management Information Systems Department.
Joe Mazur is a visiting assistant professor in the MIS Department, where he teaches courses in MIS and the economics of technology markets.
