Case study · AI, data & growth
From Averages to Per-Lead Predictions
I replaced a static return-rate average with an ML model that improved revenue forecasts, strengthened confidence in the numbers and enabled region-level bidding.
The context
Our daily revenue forecast depended on how many sold leads partners would return. We used a static average return rate, which was easy to explain but unreliable day to day because return patterns differed between weekdays and weekends and between daytime and evenings. A day could initially look strong and then deteriorate as returns arrived. We also could not close the month on the 1st because we had to wait for returns to be processed.
The challenge
I needed to make the revenue forecast more reliable while ensuring that the solution was practical to implement and trusted by leadership.
How I approached it
We initially experimented in-house with Random Forest models and several variants. The tests showed potential, but implementing and operating it well called for specialist capacity.
I therefore partnered with a specialist machine-learning company. Together, we built a data pipeline on a cloud data platform that allowed its team to run prediction models on our data. The setup progressed quickly, enabling us to move directly into tuning and improving the predictions.
Instead of applying one top-down average, the resulting model generated a return prediction for each individual lead. This created a more granular foundation for both revenue forecasting and downstream bidding decisions.
What it delivered
Forecast accuracy improved substantially, although we did not establish a precise metric for the improvement. More importantly, questions about the daily revenue figures dropped sharply, meeting the core objective of a forecast everyone could rely on. The per-lead predictions also enabled our automated bidding system to make region-level bidding decisions.
Looking back
I learned that a simple, explainable average can be useful, but it can also hide patterns that matter operationally. I would establish a clear accuracy baseline and success metric before starting; although the forecast improved substantially and questions dropped sharply, we could not quantify the gain precisely. I also learned to recognise when promising in-house experiments need specialist capacity to become reliable in practice.
The principle I carry into future roles is to match model granularity to the decision being made, while keeping implementation and trust in view. Moving from a top-down average to per-lead predictions improved forecasting and gave the bidding system a stronger basis for region-level decisions.