The other day our low temperature for the day was ~10 degrees F below our forecast low. I've noticed this before in the winter, especially when temperatures start to go below 0 degrees F (-18 degrees C). This made me wonder if the models used to forecast high and low temperature are different, and if so is one indeed more accurate? Conversely, if the same model is used, is it as accurate for both high and low temperatures?

I was able to find one student project from Dartmouth College in which a statistical analysis of forecast high and low temperatures were compared from three five years sets (1970-4, 1980-4, and 1990-4) of data from 15 different United States cities representing different influences such as latitude, water, and land influences. One of the findings of this dataset of 5,269 records was that the forecasts where more accurate in the summer than in the fall and winter. This seems to hint that warmer temperatures are forecasted more accurately than colder temperatures, at least on a seasonal basis. I'm hoping there's more information out there that can either confirm or deny this conclusion without the seasonal bias.

  • 4
    $\begingroup$ Not confident enough to be an answer: Predicting the low in winter can be problematic with calm weather. A very calm and clear night versus a bit of wind or a bit of cloudiness can easily account for that 10 degree Fahrenheit difference. $\endgroup$ Commented Dec 29, 2017 at 0:30
  • 1
    $\begingroup$ The answer to this question would depend on the forecast accuracy criteria that is currently unstated. Generally in middle latitudes the forecast accuracy depends most on alternating high and low pressure systems that depend on season, so yes that old study revealed something real. On Northern hemisphere summer the north-south temperature gradient is smaller than in winter. This temperature gradient essentially fuels these disturbances (low/high systems) so when it is smaller the disturbances are less frequent. And when they are less frequent changes are slower and more accurately predicted. $\endgroup$
    – Communisty
    Commented Dec 29, 2017 at 8:33
  • 1
    $\begingroup$ if there is inversion as in hot air on top of cold the forecast do have a too low resolution to get it right at different elevations so the temparature might be off by plus minus 10 degrees easily. $\endgroup$ Commented Dec 29, 2017 at 9:37
  • 1
    $\begingroup$ There's plenty of data out there and this would make a fascinating project. $\endgroup$
    – user967
    Commented Dec 30, 2017 at 18:36

2 Answers 2


There are two ways to answer your question: the first would be by verifying historical forecast data (calculating the difference to observed temperatures) and looking into a temperature dependency of these errors. This would certainly be doable - but in this context on Stackexchange not.

So I suppose you refer to a general, more physically interpretable explanation what the difference in forecasting higher or lower temperatures is.

Even though it makes a difference for physical processes like radiation and evaporation if temperatures are high or low, the characteristics stay the same - more explicitly said:

Temperature differences matter more than absolute temperatures!

For example:

  • The net radiation (budget) is proportional to the difference of soil and atmospheric temperature (each to the fourth power).

  • Deeper layers of soil keep the surface warm by transporting heat (also proportional to the temperature difference).

  • and so on...

But there is a significant difference in the active physical processes: At daytime (Tmax), the near surface air is (mostly) coupled to the temperature of higher layers of air by turbulence. On a sunny summerday, the surface air can not get significantly warmer than the air above because if it gets warmer, the air rises and cooler air from above cools the surface layer. In this way, the surface air temperature is 'tied down' so that it can not reach extrordinary temperatures without also heating up a say 1500 m boundary layer of air.

If the surface air is cooler than the air above it gets decoupled from the temperatures above - turbulence is reduced, and theoretically the temperatures could drop a looong way down, until the incoming (longwave) radiation from the atmosphere (say hi to greenhouse gases) stops this temperature drop. The equilibrium temperature depends mostly on the amount of water vapor above our heads -> warmer in the tropics, cooler nights in polar regions. But the equilibrium is practically never reached. In most cases it is a dynamic thing with ground heat flux and sporadic wind gusts - which are of very local effect.


Comparing these two cases one can see that the Tmax is more homogenously distributed in space because it mostly depends on the airmass. Minimum temperatures can be very different even for close locations because it really depends on your surrounding: trees, buildings, ridges. You can see the lowest Tmin at high altitude locations which are not exposed to wind gusts (sinkholes) which are covered by snow (isolated from soil heat).

So by analyzing a forecast error dataset, you would most likely see that Tmax are more accurately forecast.


The problem you may be seeing is due to how the weather industry calculates the low temperature for the day. They use the period from 7pm of the current day to 7am of the following day. Sounds completely ridiculous? Me too, you can easily observe this for yourself. Go to a site like intellicast and where it shows the 10-day weather forecast, it will also show the temperature as a graph line. You can see that the forecast lows of the day are often wrong - it is the low for the next day.

  • $\begingroup$ Weather forecasters typically use midnight one day to midnight the next day for high temperatures. High temperatures are often recorded a few hours after noon, so that time frame makes sense. Some weather forecasters use noon one day to noon the next for low temperatures. That also makes sense, as low temperatures typically occur at dawn. (Others use midnight to midnight, as that makes statistics easier to compute.) I have yet to see one that uses 7 pm to 7 am the next day. That is nonsense. $\endgroup$ Commented Apr 26, 2022 at 14:51

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.