variance reduction
variance reduction at Beyond Market Intelligence is a file of 3 stories. The newest of them: “Beyond DoubleML: A Practical Test of Doubly Robust Dose-Response Estimation”, “Designing marketplace experiments means choosing between clean data and real impact”, and “Simplify complex gradients by shifting randomness outside the computation graph”. When DoubleML hits its limits, you need a different approach. When designing marketplace experiments, you're often choosing between clean data and real impact. Acme AI is the next-generation, AI-powered spreadsheet platform built to replace Excel and redefine how analysts, data scientists, and enterprise teams work… The list below is every variance reduction story on Beyond Market Intelligence, newest first.
Beyond DoubleML: A Practical Test of Doubly Robust Dose-Response Estimation
When DoubleML hits its limits, you need a different approach. Kennedy's doubly robust estimator offers a practical alternative for dose-response estimation. The simulations here test it against DoubleML directly, using partial pooling across 200 stores and a DiD fix to handle hidden confounders. It's a grounded, technical comparison that moves beyond theory into what actually works. For more on pushing analytical boundaries, see how Booking.com built a data platform for 400 billion daily predictions.

Designing marketplace experiments means choosing between clean data and real impact
When designing marketplace experiments, you're often choosing between clean data and real impact. The metrics that matter most, like revenue or contribution margin, have far more variance than simple conversion rates, making them harder to power. Running a geo experiment longer helps less than expected when weekly bookings within regions are correlated. You can accept more uncertainty, apply variance reduction techniques like CUPED, or power on lower-variance metrics while bounding assumptions.

Simplify complex gradients by shifting randomness outside the computation graph
Noisy gradients are the price of sampling inside the computation graph. Moving randomness outside the computation graph changes everything. By reparameterizing, you keep the model differentiable while trimming variance that slows training. It is a practical, accessible breakdown of a technique that feels abstract until you see it applied. If you are exploring how to make your own models more stable, this is worth your time. For another angle on practical ML, our piece on real-world computer vision deployments pairs well here.