Bets actually get smaller after a loser, using Kelly. The Kelly Criterion tells you the best percentage of bank to bet and - presuming that the game itself doesn't change in the meantime - after a losing bet you'll be using the exact same percentage of a now-smaller bank as your new stake.

It's arguable whether Kelly staking is a good method for horse betting. It was only intended for a system where: it's a non-terminating game and both the odds on offer and the odds of the event actually occuring are explicit.

Non-terminating means you won't ever go bust. You could be down to a bank of 1c and betting 0.25c and you will still eventually make your fortune (under the terms of the game he used). With a bookie, once your bet slips below the minimum allowable stake - unless you have a lucky run of wins with your remaining pennies that bumps you back above the minimum - that's it; game over.

In his coin flip game we know that - unless the coin is flawed or the procedure of flipping compromised - there is a 50-50 chance of the event occuring. In horse racing we know that the favourite has a 30% chance of winning BUT this is the average of all favourites over time. Each individual favourite actually has a totally different chance of winning - only some of which is reflected in the market pricing. Unfortunately no-one knows what the actual chance of a horse winning is before the race, and little more of an idea after it. The market is totally wrong as often as it is far too conservative. Altogether though, it averages out to just about right (fave-longshot bias and other human betting biases notwithstanding).

So the important figure is not just what the long-time strike rate of a selection is, it's also the standard deviation. If you were to use a Kelly calculator online, you'd find that they all urge you to be conservative with your estimate of the odds. Most users arbitrarily knock off a percentage point or two, but essentially they're asking you to look at your mean (e.g. 30% of favourites) and standard deviation. That is, 95% chance that the actual chance is within 30 - (1.96 * stdev) and 30 + (1.96 * stdev). Instead of plugging 30% in the calculator you would pop in the conservative 30 - (1.96 * stdev) 'cause your estimate is wrong and will always be wrong, so at least be cautious about how wrong and in which way wrong it is. You likely wouldn't bet very often, but you also likely won't go broke.