Decoder every acronym, in plain English

How we combine the member stations into a composite. For each moment we throw out the single highest and single lowest reading, then average what's left (the middle three or more). It's the same idea as Olympic scoring dropping the top and bottom judge — robust to one bad sensor without needing to know which one is bad.

Global weather models forecast a generic point; they don't know the plateau runs cooler. We trained a lookup on five years of paired airport→mountain readings: for a given hour, wind, and sky condition, how many degrees does the mountain typically differ? Then we shift the raw forecast by that amount. The corrected line is labeled *_MSWX on the graph; it beats the raw model by about 1.2°F at short range.

The rate air temperature drops with altitude. The textbook "standard atmosphere" value is about 3.5°F per 1000 ft — go up a mountain, it gets colder. The lapse page measures the actual rate between our stations, hour by hour. When it goes negative (warmer up high), that's an inversion.

The atmosphere flipped upside-down: warm air sitting on top of cold air, the opposite of normal. On clear, calm nights the valley floor radiates its heat to space and cold air pools in the low ground, while the plateau stays in the warmer air aloft. Result: Monte Sano is warmer than downtown despite being higher. This is the single biggest reason a Huntsville forecast is wrong for the mountain — and the reason this site exists. Catalogued on the inversions page.

Instead of betting on one model, run several and average the result. No single model is reliably best days out, but their independent errors tend to cancel. Our headline pool forecast (POOL_ENSEMBLE) is the mean of the pool curves from GFS, ICON, and Open-Meteo's blend.

The pool forecast isn't a weather model — it's a physics simulation of the water. We track heat in and out four ways: convection with the air, solar gain, evaporation, and longwave radiation to the sky. It re-anchors on the real sensor reading every hour, then integrates the weather forecast forward a week. Details on the pool page.

The solar energy reaching a flat surface, in watts per square meter — the sunshine that heats the pool. We feed each model's own GHI forecast into the pool model rather than guessing it from cloud cover, because "cloudy" still lets ~40% of the sun through, not 25%.

Average size of a forecast's miss, ignoring direction. For temperature, under 2°F is genuinely good. The headline number on Verify.

The average signed error (forecast minus reality). Positive = the source runs too warm; negative = too cold. Unlike MAE, bias is fixable — it's exactly what bias correction subtracts out.

Like MAE, but it squares the errors first, so big misses hurt more than small ones. If RMSE is much larger than MAE, the source is occasionally way off rather than steadily a-little-off.