Classic Time-Series Forecasts (2008)

In 2008, Lokad was launched with a forecasting engine offered as a Software-as-a-Service (SaaS). The company’s original tagline was forecasting as a service. This engine provided classic point time-series forecasts. Over the years, Lokad matured by introducing more advanced predictive technologies and adopting a broader supply chain perspective beyond its initial demand forecasting roots. The original forecasting engine was gradually phased out in the mid-2010s and ultimately shut down in 2020.

The original engine (now defunct), introduced in 2008, operated as a meta-model that contained a suite of forecasting models—mostly of the autoregressive variety—along with a classifier used as a model selector. This selector chose the most suitable model for each time-series. From 2008 to 2012, the engine was progressively enhanced with additional models and a more refined selector.

Later enhancements included non-parametric models influenced by the machine learning approaches popular at the time. These models utilized a concurrent time-series perspective, enabling the engine to apply appropriate seasonality coefficients even for time-series that did not have a full year of historical data. Similarly, this allowed the engine to forecast product launches by leveraging similarities identified through tags assigned to the time-series.

While Lokad never used one client’s data to improve another client’s forecasts—a commitment that still holds true—there was only one forecasting engine shared among all clients, using the same meta-parameters. Consequently, Lokad had to set up high-quality default parameters that would perform well across varied scenarios. In an indirect way, the engineering insights gained from one client ended up benefiting others (and vice versa).

Ultimately, this engine was phased out as Lokad developed superior technologies. Although the model selection approach might seem like a contest of competing models, it led to substantial instability in the forecasts. Adding just one more day of historical data could prompt the selector to switch models for numerous time-series, causing erratic fluctuations. This issue is inherent to any forecasting system relying on internal competition and is now viewed as an outdated design at Lokad.

The lack of probabilistic output was the second significant problem. Point time-series forecasts completely overlook uncertainty, rendering decisions based on these forecasts highly fragile. Indeed, if actual outcomes deviate from the forecast, economic performance often degrades sharply. Lokad addressed this limitation by introducing probabilistic forecasting technology.

Finally, as a third major issue, the rigidity of a time-series-only perspective posed a serious limitation. Time-series—represented as one-dimensional vectors—offer limited expressiveness for historical data. Even in supply chain situations where point forecasts might suffice, a pure time-series framework typically fails to capture the full complexity of real-world contexts.

The classic forecasting engine was definitively phased out in 2020, following the launch of our differentiable programming technology. Although point time-series forecasts are no longer recommended, differentiable programming can readily produce both point time-series forecasts and probabilistic forecasts.