Extracting More Signal From Every Unit of Compute

Very proud of the MetaLearner team for designing, coding, and optimizing a novel feature-elimination methodology, allowing large-scale forecasting systems to identify the most predictive features across organisations with thousands of products in remarkably little time.

This matters because forecasting at scale is fundamentally a high-dimensional optimization problem.

Over the years, working with different customers, MetaLearner has built a dynamic feature store encompassing statistical characteristics, time-series transformations, operational variables and external features, significantly expanding the information available to explain and forecast almost any time series.

The challenge with richer information is that it creates a much larger computational search space. More features mean more variables to evaluate across multiple forecasting algorithms, rolling validation windows, hyperparameter configurations and ensemble combinations. Without intelligent dimensionality reduction, a significant amount of computation is repeatedly spent evaluating variables that are redundant, noisy, highly correlated, or provide little incremental predictive information.

Our newest feature-selection layer attacks that problem upstream while retaining interpretability.

Instead of asking every forecasting model to process the complete high-dimensional feature space, the algorithm identifies the small subset of features carrying the greatest predictive information and eliminates variables contributing little incremental value before the expensive model-selection and optimization stages begin.

The methodology is a high-performance extension of MetaLearner’s existing feature-elimination architecture, but with an important difference: it evaluates whether features improve forecasting out of sample across a population of time series, rather than simply how well they explain an individual series historically.

It also automatically determines how many features are worth retaining, rather than requiring us to specify a fixed feature count in advance, resolving one of the major practical limitations of conventional forward and backward Feature Elimination approaches.

The result is a dramatically smaller effective search space.

Our latest benchmark started with a small library of 132 candidate forecasting features. MetaLearner’s methodology reduced that to approximately 3 features per series, compared with approximately 9 features for conventional RFE and 8 for LASSO.

Importantly, that reduction did not sacrifice forecasting performance.

Across three independent out-of-sample benchmark samples, MetaLearner was statistically indistinguishable from the competing methodologies under the production-style XGBoost learner, while under a linear Ridge model it outperformed RFE across all three samples and was statistically significantly better in two of three, with p-values below 0.04.

When we forced the competing methodologies to use the same number of features, MetaLearner still produced the best results.

That matters because it suggests the advantage does not simply come from using fewer variables. The methodology is identifying a very small subset of features that carries a disproportionate amount of useful predictive information.

The computational implications are significant.

Reducing the model input reduced XGBoost fitting time by 64%, or approximately 2.8× faster fitting.

And this saving occurs repeatedly.

Every rolling-validation window, hyperparameter combination, candidate forecasting algorithm and ensemble component that follows now operates on a dramatically smaller feature space.

The feature-selection stage itself also scales efficiently. In our production-scale test, MetaLearner searched across approximately 6,000 time series and 132 candidate features in under 100 seconds using only four workers: equivalent to roughly 17 milliseconds of amortized wall-clock time per series.

And that benchmark was intentionally constrained to just four workers. With greater parallelism and further optimization of the feature-generation layer, the computational advantage becomes significantly larger and provides a path toward ~80% reductions at pipeline scale.

The significance goes beyond raw speed.

Because MetaLearner does not rely on a single forecasting model, these efficiencies compound. Our framework can dynamically select relevant features, incorporate related-product and exogenous information, and assign an optimized forecasting model stack at the product level. It evaluates models across both forecast error and forecast reliability rather than simply choosing one globally “best” algorithm.

In other words, we are using intelligence to reduce the amount of brute-force computation required to produce intelligence.

That is particularly important in an AI Factory architecture.

GPUs give us the ability to perform forecasting, simulation and optimization at enormous scale, but better algorithms determine how efficiently we convert that compute into useful decisions. MetaLearner’s role is increasingly to abstract the underlying computational complexity so enterprise operations teams can access these capabilities without needing to become GPU, CUDA or forecasting experts themselves.

All these leads to lower infrastructure costs and the ability to quickly deploy increasingly sophisticated decision intelligence across thousands, and potentially millions, of product-location-time-series combinations.

The objective is not to use more compute. It is to extract more intelligence from every unit of compute.