1- I was wondering if someone had made a stock market prediction model that worked, and made money by his prediction.
2-any tips about how to make a model for this problem
3- and what was your experience doing this mode?l is it realy makes money or it’s not accurate to give you even a trend for the stock prices?
Hi Bahaa!
Yes, people have done this successfully.
Focus (usually) on returns rather than asset prices.
If you know that certain asset prices are correlated (because of rigid
arbitrage-like relationships or because of “soft” statistical correlations),
preprocess your data to take advantage of those correlations. Don’t make
your network “learn” those correlations. Your correlations need not be
rock-solid – it will generally be better to ask your network to “learn” the
deviations from the baseline correlations, rather than those correlations
themselves.
Try to take into account relevant non-tradable information (i.e., things other
than prices) – for example, the “fundamentals” of companies whose stocks
you might want to trade or the observed liquidity of assets whose prices you
imagine predicting.
It’s quite hard and it’s a moving target. People are all the time trying to do
this (with or without neural networks). When some subset of these people
find a pattern that lets them make tradable and money-making predictions,
they make those trades (to make money, naturally) – and those trades push
the prices back to where the tradable pattern no longer exists.
Let’s say that both you and I build an MNIST digit classifier and yours is
quite good and mine is a little better. We both have good classifiers, so
we’re both happy. But if both you and I build a model that predicts some
stock’s price and mine is a little better, I’ll trade using my model (to make
money, naturally) and squeeze the predictability of that stock’s price out
of the market. I’ll end up with money, but you, maybe not so much …
Good luck.
K. Frank
I’m focusing on making a model that is able to to tell me if the percentage is 5% or higher.
It’s still in progress I can’t tell yet if it will work,
It was hard to predict the next day prices. so I’m to predict the percent change between the current day and the next day(future day)
Yes, ive made a few. The best one i have is a neural net i made from scratch. Try to develop features that show how the value changed instead of the actual value. Example: instead of having a feature like rsi[0] show the change like rsi[0] - rsi[1] which is the difference between the current rsi val and the previous. Scaling is probably the most crucial aside from feature selection. Ive found that when your features show how the value changed you also set yourself up to create stable scaling methods. Wherever possible try not to use cumulative features. You will consistently have scaling issues. Stay away from rolling windows if you can. Min max linear scaling has given me the best results. IQR/ZScore or what they call ‘robust’ scaling has always had a a profoundly negative impact on the predictions. What i typically do is scale all of my features and labels into positive values (0, 1) and i scale the output back into the original label range …this will provide you with both negative and positive predictions. Another important point… balance your label scaling… meaning dont scale out of a skewed range and then use the skewed values for wcaling into the new range. It makes it exponentially harder to train your model. Another VERY important part to donate your attention to is the distribution of your features. Yes you are probably using non-linear activations (although ive experimented with L. Regression and the predictions were for the most part directionally accurate) however, here you want to watch the distribution in the area right around the MAX and MIN, if thats not a heavily populated area region then pay attention to the grouping that is closest. Your looking for any feature distributions that may be converging or diverging. For instance if feature 0 is rising and feature 2 is falling paired with symptoms like what seem to be inverted predictions you should consider incompatible features. Also yes, daily closing prices would be extremely hard to predict i would be very impressed if someone were to create a reliable model that could predict them while having a mse == the daily range. I have had the most success with futures, 5 minute bars and 3 bar predictions. So essentially 15 min windows
if only we could build a predictor that pushed the price around. The sad truth is that the Market is something that takes a lot of time, years to become consistently profitable. Even with a classifier or predictor that is 95% percent accurate or has an mse of .00009. I run a couple neural networks in the futures market. About 80-90% of market volume is executed algorithmically.