Temperatures, however, are not random and probabilistic; they are the explicit result of many factors. These same factors are used to predict long-term climate change. If the climate models that use these factors work so well for the future, why do they not work well for the next week? It's the same model, the same factors. In the analogy of the lottery numbers, if you mess up the near-term probabilities, you can rest assured the error is magnified over a long term.
Consider that the ball being drawn out of the lottery machine is not truly random and probabilistic, either - it is a direct result of the laws of physics, the starting condition of the ball-drawing machine and the inputs to it.
Just like the ball-drawing machine, the short-run weather in the next minute, hour, day or week is a highly noisy process that can be accurately modelled as a random, probabilistic process. This is how weather forecasts work - they'll give the probability of rain in the next 24 hours.
Scale matters. Most climate prediction software looks at weather on a far vaster, long-term scale than hours, days, or even weeks. Variations over such short periods ultimately have little effect on long-term trends.
Hold on a moment. I can't predict tomorrow's lottery numbers with any reasonable accuracy, but I can make an accurate long-term prediction about how often a given number is going to appear over the next decade of lotteries.