Potentially almost any data input can be used as seed. But in general you don't want to use information that can be trusted to have high entropy and be secret all the time. This rules out stock exchange and the weather for instance, as the stock exchange may be down, and there is not much entropy in a clear blue skye.
Almost all of the sources you mention rely on outside phenomenons, which means you have to trust the source. Moreover, you'll need a trusted path to your source, and the trusted path likely requires a random number generator. So you would get into a chicken/egg situation.
What is usually done is to rely on a local phenomenon and then use highly unpredictive measurements. Local phenomenons can be kept secret and measuring in the extreme will mean that you aren't dependent on large fluctuations. For instance a HDD can be used to inform the system about I/O times, and measuring in nano-seconds will give enough entropy for a few bits of seed.
Of course, you neverk know what kind of system you are going to find. So fortunately it starts to be commonplace to include a random number generator in CPU's. An RNG in the CPU or - for instance - in a TPM is probably the best RNG you can find.