Automated search at scale
Automated feature evaluation, and the path to agentic model search.
With a uniform contract and a programmatic Features API in place, model and feature search stops being manual labour and becomes something you can automate. That is the part of this project that points most directly at where data-science platforms are heading.
An automated feature study
I ran an automated meta-study to measure what each exogenous feature was actually worth. Rather than eyeball a single tuned model, the study looped an Optuna hyperparameter search across every candidate feature and every demand class, then compared each feature's effect against a matched baseline.
A controlled detail made the comparison unusually clean. Because the searches shared a fixed set of starting trials, each feature was evaluated against an identical configuration, so the difference in error could be attributed to the feature itself rather than to lucky tuning. Aggregated across many such pairs, this produced a defensible read on which signals helped, which did not, and where their effect varied by demand type and over time.
There is an honest lesson in it too. The number of trials per study was initially set far too low for the size of the search space. But because every study shared the same startup trials and the same seed, that mistake became an asset: it made the paired comparisons exactly controlled. Reporting that plainly is part of the method. A platform's job is to make results legible enough that you can see both what worked and why.
Toward agentic search
The endpoint is the interesting part. Once every model speaks one contract and every feature is reachable through one API, an agent can drive the search itself: select models, assemble inputs, evaluate and tune them, semi-autonomously, in response to a prompt. Exposing the contract-compliant library and the Features API as tools an agent can call is a short step from where the platform already stands.
This is the shape of the thing. The platform is the durable asset; automated, scalable search is what it makes possible. Strong models still do the forecasting. That was the team's work, and it was substantial. But the rails they run on are what let a team, or an agent, build them efficiently, compare them fairly, and trust the result.