One (rather easy?) thing I'd like to see in weather forecast apps is a confidence score.<p>I know that sometimes the weather is hard to predict. Right now I can't tell by looking at the weather forecast how confident they are. If the different weather models give significantly different forecasts, the confidence score should reflect it.
How good is weather prediction for other people? Is it actually accurate for you? Are even the <i>current conditions</i> accurate for you? I use Weather Underground, and a few weeks ago in the Bay, here are just a couple of the things I remember observing (out of the <i>many</i> wrong predictions): (1) One afternoon I was told there would be rain in 5-6 hours, then a couple hours I was told there would be none. (2) I started getting soaked while walking outside, and I checked the <i>current conditions</i> and was informed it is in fact not raining and there is no upcoming rain either. These were despite the facts that the weather stations were < 1 mile away for me and the conditions persisted for a fair bit (i.e. it wasn't just a random 2-minute shower).
Not widely publicized is that 5G cellular is going to eliminate one of the primary data inputs that have made weather prediction successful.<p>The US FCC decided, on their own, that this was not an important problem, compared (most likely) to the amount of money to be made building out 5G.
I have to shout out for windy.com (I am not its developer sadly). It is probably the best weather site (and app) out there. You can even choose between weather models.
I often wondered the same thing myself, why don't weather providers provide their own accuracy rates? I started to make a "simple" system for estimating NOAA accuracy, and immediately ran into trouble determining what accuracy meant.<p>If they predict a high to be 86deg, and it's really 85, what does that mean as far as accuracy goes? If we use the Kelvin scale, even a 10 degree error makes it seem pretty accurate, though a person's experience in those extremes will be very different.<p>But I think the biggest problem is that the simple weather forecasts that we use on a daily basis, is a poor representation of what weather forecasters actually do. They're modeling how weather systems form, move, and interact. If a model predicts storm forming and moving a particular direction, but the 10 day forecast is off by 100 miles causing it to rain a day later, what does that mean for accuracy? Another model could just use the average weather as their forecast, and might score pretty high as far as long term accuracy, but would be pretty useless from a user's perspective.<p>So, if someone forecasts a high of 86 with a 99% confidence level. What would that mean. That it'll be 86 somewhere near there, that it'll be close to 86 at that location that day, or that it'll be 86 at that location within some timer period? You really can't boil all of those variables down into a single number.<p>And then you'll run into issues tracking the confidence of the confidence levels. Ad infinum.
NOAA hosts a Python SDK and REST API for historical data. And I believe they present at the SciPy conf every year (upcoming in July)<p>I found this talk by Uber's Danny Yuan super insightful. Forecasting is probably the subset of ML I am most excited about ;)<p>Two Effective Algorithms for Time Series Forecasting<p><a href="https://www.youtube.com/watch?v=VYpAodcdFfA" rel="nofollow">https://www.youtube.com/watch?v=VYpAodcdFfA</a>
It's kinda cool that we <i>can</i> predict in enough time to save lives and equipment. If our brains or metabolism had been slower, then this might not be the case, even after we've developed the mathematical theories that we have now.