Skip to content
🧪experiment

Forecasting Under Fire, The Experiment

  • forecasting
  • bayesian
  • experiments
  • oil
Forecasting Under Fire
Part 1 of 12
0:000:00

Forecasting Under Fire, The Experiment

The war started three weeks ago. Since then, I've refreshed the news more times than I can count, learned nothing useful, and felt progressively worse about everything.

I needed to do something with the helplessness. I had to stop refreshing the news every five minutes. But the forecasts were bugging me too. They didn't seem to match up with reality. Goldman Sachs changed their forecast four times in six days.

So for the next eight weeks, I'm running an experiment. Weekly oil price forecasts, published every Sunday. Not because I'm an expert. Because trying this systematically, as a person, beats doom scrolling my way through a crisis I can't control.

I need to say this clearly. This experiment is about processing uncertainty and learning to update beliefs under pressure. It is not about treating a war as an intellectual exercise. Our hearts are with the innocent people across the region. This needs to end. Forecasting oil prices doesn't fix that. But it does give me something structured to do with the helplessness instead of refreshing my phone every five minutes.

The Experiment

Every Sunday through 10 May, I'll publish a point estimate for Brent crude seven days out, confidence intervals at 68% and 95%, the scenario probabilities that generated them, and what I got wrong the previous week.

The method is scenario-weighted forecasting. I assign probabilities to five scenarios, each with price targets and volatility assumptions. The forecast is the probability-weighted outcome.

Week one probabilities:

  • Escalation (45%). War expands, Strait of Hormuz closes for more than three weeks
  • Stalemate (25%). Prolonged conflict, intermittent disruption
  • De-escalation (15%). Ceasefire reached, gradual normalisation
  • Demand Destruction (10%). Economic slowdown reduces consumption
  • Black Swan (5%). Something I'm not seeing yet

These will change every week as new information arrives. The point is to learn how to update beliefs when events contradict your priors, not to nail the prediction on week one.

This Week's Forecast

Brent crude, targeting 20 March 2026

  • Point Estimate: $120.16
  • 68% Confidence Interval: $85.11 - $138.08
  • 90% Confidence Interval: $71.21 - $145.15

Current price (13 March close): $98.91

The wide intervals reflect genuine uncertainty. If you're looking for precision, you're reading the wrong forecaster.

Why Oil

Oil prices move faster than any other economic signal during geopolitical crises. They respond to news within hours. By the time bond yields shift or unemployment data arrives, oil has already told you what's coming.

It's also measurable. You know when you're wrong. No room to fudge the outcome.

Why Me

I'm not an oil analyst. I'm not a geopolitical expert. I have no edge on information.

What I do have is a track record of spotting pattern shifts before institutions admit they're happening. And a growing suspicion that institutional forecasting is optimised for reputation management, not truth.

I'm not always right. But when I'm wrong, I know why.

The Rules

  1. Forecasts locked in every Sunday
  2. No edits after publishing
  3. Full accounting of errors every week
  4. If I score worse than a naive baseline (last week's price plus random walk), I'll say so
  5. Everything is public. Code, data, methodology, and a live dashboard that updates in real time

What I'm Actually Testing

Can I learn to update probabilities cleanly when new information contradicts my priors? Do I overweight recent shocks or underweight long-term fundamentals? How long does it take me to recognise when a scenario is dead?

Eight weeks is enough time to be wrong in interesting ways. Not enough time to get comfortable.

Follow Along

The data lives on GitHub. The code is open. The dashboard shows forecasts, scenario evolution, and accuracy tracking as the weeks progress. If you want to run the same model with different probabilities, go ahead. If you think my scenario definitions are wrong, show me yours.

I'll be updating this every Sunday through 10 May. By then we'll know if this was useful or just public accountability for a failed experiment.

Either way, it'll be documented.

Methodology Notes

  • Data source. Yahoo Finance (Brent: BZ=F, WTI: CL=F), fetched daily at 02:00 UTC
  • Forecast horizon. 5 days
  • Update frequency. Weekly (Sunday)
  • Baseline comparison. Naive forecast (current price as prediction)
  • Scoring metric. Mean Absolute Error (MAE), calibration plots for confidence intervals

Next update. Week 1, 15 March 2026

Notes