"News" is the difference between what happens and what you expected to happen. If you have rational, and hence unbiased, expectations, then the news, on average, should be neither good nor bad. The good news and the bad news should cancel out.
So why does the news that gets reported seem mostly bad news? Does bad news sell more newspapers than good news? Are newspapers biased?
Suppose reality is skewed. The distribution of the news has a zero mean by assumption, if our expectations are unbiased. But suppose the distribution has a long tail on the bad news side, and a short tail on the good news side. And suppose newspapers only report on big news, that's (say) one standard deviation or more away from the mean. Then most of the news reported would be bad news, because there are very few good news stories that are big enough to be worth reporting, just lots and lots of little good news events. "There were no serious accidents today on highway 401" never makes it into the newspapers.
But why should reality be skewed? That's the tricky question. But I think I can answer it.
(I'm not sure if this answer is original. I apologise for my inability to explain it clearly in math. This all came out of a lunchtime conversation with Frances, where she was trying to think of the best and worst news story of 2011.)
Suppose first that reality were the sum of the independent actions of a large number of beings who didn't pay any attention to what humans wanted. Reality would then be normally distributed. The distribution of news would be symmetrically distributed with mean zero. Newspapers that reported only one standard deviation events would then print an equal number of good news and bad news stories. The weather is a bit like that.
Now let's introduce a human optimiser. Let there exist a social welfare function W(R), where R is reality, and W(R) is human welfare given that reality. Let reality at time t be the sum of human choice H plus random noise V(t): R(t)=H+V(t) where V(t) is normally distributed with mean zero. The human optimiser chooses H to maximise the expectation of W(R(t)). So H* = argmaxE(W(H+V(t))).
If W(R) is a strictly concave function, that is symmetric around its maximum value at W(H*), then the distribution of the news W(H*+V(t)) around the mean E(W(H*+V(t))) will be skewed. A newspaper that reported only one standard deviation events would mostly report bad news. Most of the time, the news is good news, because V(t) is mostly a small number, but it's not good enough news to be worth reporting. When V(t) is a large number, either positive or negative, then it's very bad news, and worth reporting.
This seems correct to me, but I can't do the math. Somebody reading this ought to be able to do the math. The best possible news is that V(t)=0, which is also the mode of the distribution, and so human welfare at V(t)=0 is only a little bit higher than mean human welfare, and not worth reporting. The worst possible news is some large value for V(t), either positive or negative, which is a rare event, but very bad news and worth reporting.
Here's an example.
Suppose I'm driving a new car 100 kms to Maniwaki. Since it's a new car, and the product of a human designer, it will almost certainly get me to Maniwaki without breaking down. If it has a 99% probability of getting me to Maniwaki, and a 1% probability of breaking down, anywhere along the road, the expected distance it will take me is 99.5 kms, which is the mean of the distribution of reality. The distribution of reality around that mean is highly skewed. There's a very big spike at 100kms, and a long thin left tail all the way down to 0kms. If it gets me to Maniwaki, that's good news, but not big enough news to report. 100kms is too close to the mean 99.5 kms. The only news worth reporting is if it broke down. Which is bad news.
As the car gets older, and less a product of human design and more a product of a nature which cares nothing for my wants, the chances of it breaking down gets bigger, and the distribution of news becomes less skewed. If it's a very old car, and very unreliable, it would now be big news if it actually got me to Maniwaki.
Most of the bad news that gets reported is human news. It's human news that has a skewed distribution. The weather is not a product of human optimisation, so news about good weather is as likely as news about bad weather.
The optimiser needn't be human. It could be some evolutionary process. Most big genetic mutations that are worth reporting are bad news for the organism. (I read that somewhere). [Update 3: read the first comment by Jeremy Fox on this. Jeremy understands this stuff, and explains it clearly.]
Update 1: my comment in response to Min and Frances may help with the intuition:
Suppose (OK, no need to suppose?) I were a small c conservative, who didn't like any change in any direction. Then a day when there was no change would be good news. If change were a mean zero normal distribution, then changes to my utility would have a skewed distribution. Lots of days when all changes are small, so I am slightly happier than I expected to be, but it's less than 1 SD, so it's not worth reporting. And a small number of days when there's a big change, and my utility is more than 1 SD lower than i expected, so it gets reported.
Now, any designed system, like a new car, is very much like a small c conservative. Since the designers and builders (presumably) optimised all the settings originally, any change is a bad thing. (If a change made the car go faster, or use less gas, or corner better, the designers would have built it that way to begin with.) So, if all changes to the car are normally distributed, the probability distribution of my utility will be skewed.
My math model was trying to formalise some complex designed system, like a car. H* is the original manufacturer's setting of the car's mechanicals. V(t) is all the things that wear and tear and nature does to the car's mechanicals. Even if V(t) is normally distributed, W(H*+V(t)) will have a skewed distribution.
Udate 2. Scott Sumner says in comments: "This reminds me a bit of the asymmetry in business cycles."
I wish I had thought of that. It's like Milton Friedman's "plucking model" (I think).