BLOG

Research & Strategy

Deep dives on weather market mechanics, forecast blending, and risk management for algorithmic trading.

Subscribe via RSS

GFS vs ECMWF: What I Learned Blending Forecast Models for Live Trading

Everyone has an opinion about which weather model is better. After a year of blending them for live trading, here's what the data actually says—and what it doesn't.

weather forecastingGFSECMWFensemble modelsprediction

There's this old saying in meteorology: "GFS is for the US, ECMWF is for everything else."


It's not quite that simple. But it's not entirely wrong either.


The Setup


When we started building Weatherman Trading, we knew we needed forecast inputs. Two choices:


  • GFS (Global Forecast System) — US National Weather Service. Runs 4 times a day, ~0.25° resolution, ensemble of 21 members. You can get it free. It's over-fit to US conditions.

  • ECMWF (European Center for Medium-Range Weather Forecasts) — The Europeans' answer to "what if we spent more money and hired smarter people." Runs 2 times a day, ~0.1° resolution, ensemble of 51 members. More expensive. Generally considered more skillful.

  • The naive approach: use ECMWF because it's better.


    The real approach: use both, blend them, and let the data tell you what each model is actually good at.


    What Blending Means


    We pull the ensemble mean from each model for each contract location and lead time. For a high-temperature prediction in Phoenix on day 3:

  • GFS ensemble mean: 41°C

  • ECMWF ensemble mean: 39°C

  • Our blended forecast: weighted average of both, optimized via regression on historical verifications.


    The weight isn't fixed. It depends on:

  • Location (GFS is better over the US interior)

  • Season (ECMWF tends to handle El Niño transitions better)

  • Lead time (different models have different skill decay rates)

  • Variable (temperature vs precipitation vs wind)

  • So the blend is not a fixed 50/50 split. It shifts based on those factors, and the fitted weights vary enough across markets and seasons that publishing an example would be misleading anyway.


    What a Year of Data Tells You


    After ~250 days of live trading:


    GFS is better for:

  • Short-range forecasts (0-3 days)

  • Temperature predictions in the continental US

  • Cases where you have recent surface observations

  • ECMWF is better for:

  • Medium-range forecasts (3-10 days)

  • Precipitation forecasts globally

  • Capturing low-pressure system evolution

  • Cases where pattern recognition matters more than local detail

  • Surprising findings:

  • The ensemble spread (uncertainty) is often more useful than the mean forecast. A wide spread means neither model is confident. Trade smaller. A tight spread means both models agree. That's where the edge actually is.

  • GFS has gotten noticeably better in the last year. Whatever the National Weather Service changed in their assimilation cycles is working.

  • Precipitation forecasting is fundamentally harder than temperature. Both models are mediocre. Don't overweight those contracts.

  • The official skill scores (RMSE, bias) correlate poorly with trading profitability. A model can be "skillful" in traditional metrics and still lose money if its errors are biased toward the contract strike prices.

  • The Uncomfortable Truth


    You can spend a lot of time optimizing the blend. You can fit separate weights for each season, location, lead time, and variable. You can build machine learning models to weight the forecasts dynamically.


    And maybe you get a 2-3% improvement in RMSE.


    But trading profitability doesn't scale linearly with forecast skill. A 2% better forecast doesn't mean 2% better returns. It might mean 0.5% better returns, or 5% better returns, or no difference at all—because the limiting factor might not be forecast accuracy. It might be sizing, or fees, or execution slippage.


    So here's what we actually did:


  • Fit a simple regression on 2 years of historical data: verification ~ GFS + ECMWF

  • Got weights. Used them.

  • Tested on out-of-sample data.

  • If that was profitable, shipped it.

  • Didn't optimize further.

  • Why? Because over-fitting a forecast blend to historical data is one of the easiest ways to build a system that looks great in backtests and collapses in live trading.


    The Divergence Veto


    Here's where it gets interesting for trading:


    When GFS and ECMWF disagree sharply, we take that as a signal that the forecast is uncertain. We shrink position size proportionally to that disagreement.


    If the two models produce meaningfully different predictions for the same location and time, neither one is confident. So we trade the contract at a fraction of normal size — or don't trade it at all. The specific thresholds and shrink factors are part of the calibrated system.


    This rule probably costs us more in missed profit than it gains us in avoided losses. But it's paid for itself in risk management. When both models agree, you can trade bigger. When they disagree, you have to be humble.


    Why This Matters


    The weather is deterministic at large scales (atmospheric fluid dynamics, energy balance) and chaotic at small scales (convection, turbulence). Forecast models work really well for 3-5 days, then skill degrades rapidly as uncertainty grows.


    If you're building a system that trades weather contracts, you need to understand that skill profile. You need to know what each model is good at. And you need to be willing to trade smaller (or not at all) when the models disagree.


    You can't just average two forecasts and pretend you've solved the problem. The useful signal is in the disagreement itself.


    Blend the models. Compare them. Let the divergence inform your sizing. That's where the real edge is.