Lambcast

Energy Economics • AI Infrastructure • Applied Econometrics

Forecasting Practice

In early 2026 I entered the Metaculus Cup as a way to build a structured forecasting practice alongside my research. The two questions below resolved in April 2026. I got one wrong in a way worth explaining.

What I learned from this: forecasting as a practice rewards something different from what I expected. The skill is real, but it accrues slowly over hundreds of questions and depends almost entirely on coming back to update every time new information arrives. It's less like research — where depth of analysis compounds — and more like a sport where the training is the thing. I'm not abandoning it, but it's not the center of this site. The methodology work that came out of building these models was worth doing regardless of the outcomes.

Resolved Forecasts

Silver Price: Highest Price per Troy Ounce in April 2026

Submitted: Feb 28, 2026  •  Resolved: April 30, 2026  •  Platform: Metaculus
My Median
$115
80% Range
$105 – $126
Resolved At
[FILL IN]

I built a Monte Carlo model (10,000 paths, gold β=2.29, R²=0.72) calibrated on the structural case for silver in early 2026: US Critical Minerals designation, China export controls, and AI/solar demand converging. The model assigned only ~10% probability to the April high staying below $97. It was wrong.

What happened: [Fill in the actual resolution and what the model missed — tariff reversal, macro shift, whatever drove the outcome. One honest paragraph is enough.]

Full methodology →

Who Will Win the 2026 FIDE Candidates Tournament?

Submitted: Feb 28, 2026  •  Resolved: April 16, 2026  •  Platform: Metaculus
My Pick
Caruana
My Probability
26%
Winner
[FILL IN]

I weighted recent classical form (30%) and head-to-head record against the field (30%) over raw Elo. The biggest departures from the community: Nakamura down 10 points for 8 months of inactivity against weak opposition, Praggnanandhaa up 6 as the dominant classical performer of 2025.

What happened: [One sentence on the result and whether the model's logic held even if the pick was wrong.]

Full methodology →