Cracking the Code: Global Models in Intermittent Time Series Forecasting

Forecasting intermittent time series is no small feat, especially when zeros litter the data landscape. Supply chains depend on precision here, as getting it wrong means stockouts or excess inventory. Traditionally, these time series have been tackled one by one with local models. But a shift is underway. Global models, fueled by neural networks and gradient boosted trees, are on the rise.

Global Models vs. Local Models

A recent study pits state-of-the-art probabilistic local and global models against each other, with over 40,000 real-world time series from five datasets as the battleground. The results are telling. Among global models, TiDE, a straightforward neural network architecture, takes the crown. Not only does it boast superior accuracy, but it also leaves local models in the dust with its lower computational demands.

The shift from local to global isn't just a trend. it's a necessity. Local models are labor-intensive and computationally hungry. They remind me of the days when every model needed its own hardware. Now, the global approach is akin to cloud computing, efficient and scalable. But let's not get carried away. TiDE might be simple and effective, but it's the exception, not the rule. Large global models, while promising, often falter under their own complexity and computational heft.

Distribution Heads and Their Impact

Among the distribution heads tested, negative binomial, hurdle-shifted negative binomial, and Tweedie, it's the Tweedie that shines, particularly in estimating the highest quantiles. It's the first time the latter two have been paired with neural networks, and the results speak volumes. When precision at the extremes is needed, Tweedie's the one to watch.

Why should we care? Because the ability to forecast accurately at the highest quantiles impacts everything from safety stock levels to financial forecasting. If your AI can hold a wallet, who's writing the risk model? This is where the rubber meets the road. And for the naysayers who doubt the shift to global models, the evidence is clear. The intersection of local and global is real. Ninety percent of the projects aren't, but the ones that are will redefine forecasting.

The Road Ahead

Where does this leave the field? It's a wake-up call. Slapping a model on a GPU rental isn't a convergence thesis. It's a reminder that while large-scale models hold promise, simplicity often wins the day. For those in the trenches of supply chain management, the implications are immediate. Adopt global models, but do so wisely.

In a world awash with data, the path forward isn't more complexity. It's smart complexity. As global models continue to evolve, those who adapt will lead. The rest will be left clutching their outdated local forecasts, wondering where it all went wrong.