So what’s wrong with this reasoning:
You think you’re probably right (otherwise you would change your mind).

Okay…

Therefore people with opinions similar to yours are more likely right than people with very different opinions.

Statement 2 should start with “Therefore YOU THINK people with…

You think you’re probably right (otherwise you would change your mind). Therefore people with opinions similar to yours are more likely right than people with very different opinions. People’s accuracy on various areas tends to be correlated; being right about one thing makes it more likely that they will be right about other things. Hence people with opinions near yours tend to be more accurate in general than the average person. But opinions from people who are more accurate should be respected. Hence hearing that people who are more accurate than average also have opinions near yours should give you increased confidence in your own accuracy.

Seems like each step is OK individually. Maybe it doesn’t quite work overall, but I could imagine something like this going on in people’s heads.

Market odds at any point in time are a good estimate of true odds. The fact that the closing line is more accurate than earlier in the day is yet another point in favor of the market’s efficiency. But since we can’t travel backwards in time, it doesn’t help much with our betting.

“No new information should shift the odds that much, so the fluctuation must have been due to punter ignorance.”

Weather, injuries, fixes, etc. are all possibilities.

I’m not saying there aren’t edges to be found – there obviously are. Nobody’s arguing that markets are perfect. But they do a very good job of aggregating information and making accurate predictions.

If I told you how to get those, everyone here would be rushing off to Betfair to bet on the horse races, and there would soon be no profit left for me (But see below)

2. Fluctuations don’t imply inefficient markets. Actually all else being equal, the opposite is true.

The factors influencing a horse race shouldn’t change that much close to a race. See my example where I got a big winner last year. You can get an idea of the ‘true odds’ by comparing the odds at ‘market close’ with the odds say two hours beforehand. The horse (Maldivian) was paying still $17 only some hours beforehand, I took a huge fixed odds contract at this price, then the price of the horse had suddenly plunged to $10 by race time! No new information should shift the odds that much, so the fluctuation must have been due to punter ignorance.

2. Fluctuations don’t imply inefficient markets. Actually all else being equal, the opposite is true.

It’s laughable how blatantly inaccurate ‘market odds’ are at gambling. You simply compare the true odds with the market odds at ‘Betfair’ and capitalize on the fluctuations. I’ve seen the odds of horses in big races fluctuate all over the place just hours before a race(typically there are fluctuations or +/- 30% above/below ‘true odds’!)

You can actually see the echo chamber in action. What often happens is that one particular horse starts getting hyped in the racing press, and this distorts the odds of the rest of the field.

Example: In last years’ spring horse racing carnival in Melbourne, a horse called ‘Maldivian’ was running in the Cox plate. Its true odds of winning were about 10-1, yet the Betfair ‘market odds’ on offer were a whooping 17-1 just two days beforehand. I unloaded my money on it, and…

Maldivian Wins
“The right time came at 4.05pm and Maldivian was sent out at $11 after being backed in from $17 since Tuesday and he gave Kavanagh and jockey Michael Rodd little cause for concern.”

I enjoyed my free trip to Europe from that one

As to you Bayesian fan-boys, again, it’s another case of a big ‘echo-chamber’ here, with each man repeating Yudkowsky’s mantra that ‘Bayes is the secret to the universe’, and the echos reinforcing each mans blind faith…but… I’ve got a Ruby program. There are three main ‘logic’ classes in my program. (Deuducer, Predictor, and Mapper). The Predictor class has all the Bayesian stuff, and it’s a ‘wrapper’ class for Deducer class. But then I found that Mapper class wraps Predictor class, which tells me that the stuff in the former can’t be reduced to the stuff in the latter. Bayesian fan boys have been warned.