Hey Guys and Gals, I am wondering whether a trading model is good or not. If a model can tell whether tomorrow's FX Spot price will go up or down by an accuracy of 65%, is it considered as good?

Could be great, could the a real loser. It depends on how much the market goes up/down versus being right/wrong. If 65% of the time, you make $1 for predicting the right direction, and 35% of the time, you lose $3 when the system is wrong, then the system sucks.

If it is true "on average" I would say it's really good and should make money (but this will be a highly volatile strategy with possibly long stretches of loser/winners).

only tomorrow up or down is not enough.

jurowilmott1: Yes I run the model continuously for the past year. The FX instrument is USD/JPYriccardo24: Yes I agree with you but don't you think predicting the spot rate in two month's time is a bit pointless since it is so far away? Also economic and fundamental factors may change any time and the fund will suffer long stretches of loss as suggested by jurowilmott1?

Why don't you tell us what it exactly predicts? Tomorrow's spot will go up or down compared to what "open"? If your prediction says tomorrow the spot price at 2pm will be up/down compared to spot price at 7am (tomorrow), then you just backtest this for as long as you have data and I would be very surprised if it doesn't make money with average 65% accuracy (but still you have to be prepared to take many losses in a row).

Calculating the Sharpe ratio of a strategy based upon your model based on the accuracy should be possible given a few assumptions on the distribution, but as always the devil lies in the tails...QuoteOriginally posted by: kennethlouHey Guys and Gals, I am wondering whether a trading model is good or not. If a model can tell whether tomorrow's FX Spot price will go up or down by an accuracy of 65%, is it considered as good?

That is the daily FX rate. I assume that is the closing rate at midnight in the UK.

I'm not talking about prediction of longer periods.65% for tomorrow is a very high accuracy.have in mind for a comparison that if you knew in advance USA Dow close to predict the next day europe DAXyou can achieve about 75% accuracy.so the model is a good input for a further trading strategy but I have some doubt that alone can be a good trading system, try to backtest it and you'll find out.I remember once when I implemented a neural network (although very simple) and I was amazed for next day prediction, but the P/L is different. large closed and not closed losses, etc.it's not possible to judge a model only from the the average.

Almost all indices have correlation with the DJIA and if the Dow goes up (especially if it goes up a lot) tomorrow then all other world indices will go up as well. Same for going down so I am not that surprise by the 75% accuracy. For my stuff at the moment I consider one FX instrument onlyI know it is crazy but I don't have a stop loss at the moment. Yeah still on the way to change my code and implement one

it's what I'm saying 65% is not soo far from the good 75%.but it's not enough. good luck

It depends on the risk/reward ratio and the backtesting results.I'd be happy with an algorithm with very low accuracy if I could prove that the rewards for success were 1000 times greater than the penalty for failure.

one year is by far not enough. you have to back-test it properly. In general, you can tune up and improve your alg to have 99% power prediction on the past data. The point is how well it is gonna predict the future data. And right, knowing the direction does not mean profit.

How many data points do you have for the last year? Are you using intraday prices? If so then a small time window is not a problem, as long as you accept your rules may need more frequent optimisation than if you had used a longer time frame. (ie Your trading strategy might have a shorter shelf-life).