A trading program that modifies its own risk management?

Question

The idea is that the machine would learn what level to put stops at, and where to put take profits. For example, the machine gets input about security ABC and learns the best place to put a stop. It then modifies this stop, based on the risk (volatility maybe?) of other positions it is taking.

The point is that a machine learning algorithm would analyze its risk and learn where to put stops based on its strategy (any sort of basic strategy, this not what my question is about). It would find that it, for example, has an expectancy of +2% per trade that it takes. Unfortunately, it also learns that it can only reach such a high expectancy when it places its stops and take profits very far from the current price. This makes the trades take significantly longer than is optimal. The machine would then modify its risk management to optimize its expectancy to the optimal amount of time that it is holding positions. It is trying to find the best way to place stops to have a short position holding time, while still trying to achieve a high expectancy.

The formula to evaluate the success of an expectancy on certain holding period would be something like this:

Expectancy / Average_Holding_Period = Evaluated_Expectancy

Where the program is attempting to optimize Evaluated_Expectancy

Does a system exist that works like this? If not, would it be easy to create, and would it be effective?

Answer 1

You are assuming that the "machine learning" correction will occur before a total wipeout, but you have not proven that this is how it will happen. The risk of total failure remains the same as with a 'non-modified' Martingale system.

You say in a comment "okay, but total wipeout before losses are recouped is very unlikely" - by the same token, profits from such a system would also be small. Go back to the gambler's roulette wheel and look at the core problem of a Martingale system [for ease of math let's assume no green 0's, and thus each bet is perfectly even odds]:

1. Bet $10 on red, and you have a 50% chance of winning $20, and a 50% chance of losing everything.

2. Then if you lose, bet $20: 50% chance of getting $40 [after your first loss of $10, and with your new $20 bet, this would be a net win of $10], and a 50% chance of losing everything.

3. Then if you lose that, bet $40, then bet $80, then bet $160, etc., every time doubling your bet so that at each moment, you have a 50% chance of winning a net $10 based on the value of your original bet in step 1.

The problem is that very soon, if you are incredibly unlucky, you could be $1,000's "in the hole", and you would need to risk the same amount of money just to come back with the same net profit of $10. Probability of being wiped out would be quite low - but the impact would be catastrophic.

It is difficult for humans to 'intuit' what the chance of an "unlikely" risk actually is - typically we will discount an "unlikely" event as "near-impossible", because large numbers can be hard to wrap our heads around. If you had a $160k line of credit, and you followed the Martingale system until you maxed out your credit, you would "win" the net $10 payout 32,767 times out of 32,768 times. ie: you would only lose 0.003% of the time. Sounds impossibly unlikely, right?

Well if everyone in the US [let's say 350M people] did this, then 10,681 people would lose their houses, and everyone else would win $10. Yes the chance is low - but the results are disastrous. Wrapping this system in the idea that "machine learning tweaking" will save you, might just be an attempted veneer to justify that the Martingale itself is an acceptable risk.