There are data pipelines, periodic deployment training schedules (hourly, daily, weekly), trend indicator models (decide whether the price will go up or down), sentiment models (external factors), and last but not least a solution to apply sentiment weights to influence predicted prices. All these together may help forecast prices for a specific domain.
However if you use models like the statsmodels SARIMA in python which they allow you to indicate the order of differencing, then price as inputs would be still a valid approach. Right ?
if you do not directly use price and instead use Fourier features derived from price- e.g., sin(Dense(prices))- then does the result become stationary? My inspirations for this question are Kazemi, Goel and Eghbali et al., 2019, and Tancik, Mildenhall, and Srinivasan, et al., 2020. Also, could you discuss standardizing the input to a z-score using the rolling mean and standard deviation over time to minimize distributional shift rather than using fixed standardization factors? This is sort of like augmenting the model with a non-parametric convolution over the input features. It’s simple to do and I think for most applications any computational expense associated with a moving average might be acceptable.
Part of the reason Fourier features are so powerful in the context of sequence modeling is that their inner products can form a low rank approximation for the “distances” between pairs of continuous values and their linear combinations, as proposed by Liutkus and Cifka et al., in their relative positional encoding work. So my intuition is that one might sort of be “internally differentiating” input signals without actually having to go to the trouble of doing anything other than learning some frequencies and phase shifts to model incremental price changes, which, while still subject to _second_ order changes in the market regime like the rate of inflation, still seems to be stationary in the first order to me.
Would it make more sense to predict the price using percentage changes? e.g. day 1 value is 0%, day 2 value is 15% instead of day 1 value being $1 and day 2 value being $1.15
Thanks for your inquiry. You have the right idea but one part is still wrong (you don't want to predict price). This is covered in my Time Series Analysis course so check that out for full details!
I have subscribed to one of your courses - TIME SERIES ANALYSIS, FORECASTING, AND MACHINE LEARNING IN PYTHON, but still not find the technique to solve the data not stationary.
What if past prices are scaled by the lastest price? It would keep the psst and predicted price in the same scaled, avoiding extrapolation, as well as keeping the same range of values despite not being stationary.
I guess no lag only method is actually of much use and if someone finds on that could be, won't be telling everyone... a quant trading framework doesn't need to be actually accurate to predict the price or return to profitable, sometimes slightly better than guessing can do some good.
Not roasting courses, providing mathematical evidence against a flawed method, which appears not just in courses but blogs, tutorials, and on Github. Watch the video again and you will see that I've not roasted any courses. ;)
Just anwering your dumb question, even if you developed a very good trading strategy, to open an account say in Interactive Brokers and maintain regularly trading margins will deplete your bank accounts until you become a pro, it's a big investment, and also for any Day Trading Robot development you need to be a real pro developer to account and handle microsecond Tick Data, so developing some trading strategy is very far and deifferent than trading with -> you need a lot of trading experience and huge capital. As of ny comment before, in my opinion, you are exactly the same as your critic -> you are interested in making videos and selling your stuff since if you could you will make one 2 hour video covering all that you say properly, but... fakes don't do that, not lazy probably incompetent.
This is called a "strawman argument". Look it up. Nothing you've said is even remotely close to what I am saying. Stop inventing random things out of thin air.
@@abertj.7365 This video is about math, not trading. Do you know the difference? You can't even identify the *subject* of the video, never mind being able to intelligently discuss its contents. Stop embarrassing yourself.
@@LazyProgrammerOfficial So now you want to divert your laziness and say you preach about the correct math to one who holds Ph.D. in Math? Lazy one? Stop embarrassing yourself, you know nothing about math or trading, maybe you know some arithmetics and have seen some Hollywood movies about trading.
Sadly your comment is incomprehensible, can you try rephrasing? If you are sure of yourself, then please continue the discussion and tell me why I am wrong.
@@LazyProgrammerOfficial This is not incomprehensible. You mentioned about another course isn't good at all, but after watch your video, you didn't show any improvement, or briliant approach to resolve the problem. If you're kind of person who likes to be rewarded and praised then just delete my comment.
@@anhcoder Your thinking is incorrect (logical fallacy) and it's a problem which pervades research too. Showing that something is factually wrong doesn't require you to come up with a "better approach". A null result is a valid result. For instance, showing that stock prices closely follow a random walk indicates what further approaches should and should not be attempted. > If you're kind of person who likes to be rewarded and praised Sounds like you are taking the video as a personal attack (maybe you took this approach in the past?) instead of thinking scientifically.
There are data pipelines, periodic deployment training schedules (hourly, daily, weekly), trend indicator models (decide whether the price will go up or down), sentiment models (external factors), and last but not least a solution to apply sentiment weights to influence predicted prices. All these together may help forecast prices for a specific domain.
Meucci covers this in great detail in "Risk and Asset Allocation" - the first step is always setting up a stationary domain.
However if you use models like the statsmodels SARIMA in python which they allow you to indicate the order of differencing, then price as inputs would be still a valid approach. Right ?
if you do not directly use price and instead use Fourier features derived from price- e.g., sin(Dense(prices))- then does the result become stationary? My inspirations for this question are Kazemi, Goel and Eghbali et al., 2019, and Tancik, Mildenhall, and Srinivasan, et al., 2020. Also, could you discuss standardizing the input to a z-score using the rolling mean and standard deviation over time to minimize distributional shift rather than using fixed standardization factors? This is sort of like augmenting the model with a non-parametric convolution over the input features. It’s simple to do and I think for most applications any computational expense associated with a moving average might be acceptable.
Part of the reason Fourier features are so powerful in the context of sequence modeling is that their inner products can form a low rank approximation for the “distances” between pairs of continuous values and their linear combinations, as proposed by Liutkus and Cifka et al., in their relative positional encoding work. So my intuition is that one might sort of be “internally differentiating” input signals without actually having to go to the trouble of doing anything other than learning some frequencies and phase shifts to model incremental price changes, which, while still subject to _second_ order changes in the market regime like the rate of inflation, still seems to be stationary in the first order to me.
I'd this problem 2 years ago when I used LSTM and arima to do urea price forecasting...it's great to see a video that confirms my results too..
Would it make more sense to predict the price using percentage changes? e.g. day 1 value is 0%, day 2 value is 15% instead of day 1 value being $1 and day 2 value being $1.15
Thanks for your inquiry. You have the right idea but one part is still wrong (you don't want to predict price). This is covered in my Time Series Analysis course so check that out for full details!
I have subscribed to one of your courses - TIME SERIES ANALYSIS, FORECASTING, AND MACHINE LEARNING IN PYTHON, but still not find the technique to solve the data not stationary.
What if past prices are scaled by the lastest price? It would keep the psst and predicted price in the same scaled, avoiding extrapolation, as well as keeping the same range of values despite not being stationary.
Yes, that is one way I have seen. Unfortunately, it appeared to be no more successful than any other lag-based methods.
I guess no lag only method is actually of much use and if someone finds on that could be, won't be telling everyone... a quant trading framework doesn't need to be actually accurate to predict the price or return to profitable, sometimes slightly better than guessing can do some good.
When will there be a udemy discount on timeseries and the finance course?
The best value (amount of content per dollar) is available at Deep Learning Courses (link in description).
Bro roasted courses and then proceeded to market his own course 😳
Not roasting courses, providing mathematical evidence against a flawed method, which appears not just in courses but blogs, tutorials, and on Github. Watch the video again and you will see that I've not roasted any courses. ;)
@@LazyProgrammerOfficial oh, now rewatching it sure makes sense. I beg your pardon sir, it was my fault for not understanding what you said.
Hi, where is your video about why min-max scaling is bad?
czcams.com/video/Vfx1L2jh2Ng/video.html
@@LazyProgrammerOfficial Thank you!
Can I get the coupon now?
Thanks for your inquiry. All coupons can be found at lazyprogrammer.me. Best regards
@@LazyProgrammerOfficial Thanks for creating such amazing contents!!
@@LazyProgrammerOfficial can you please share any new coupon for it
@@AhmedMohammed-wq1or Coupons are available via my newsletter
Just anwering your dumb question, even if you developed a very good trading strategy, to open an account say in Interactive Brokers and maintain regularly trading margins will deplete your bank accounts until you become a pro, it's a big investment, and also for any Day Trading Robot development you need to be a real pro developer to account and handle microsecond Tick Data, so developing some trading strategy is very far and deifferent than trading with -> you need a lot of trading experience and huge capital. As of ny comment before, in my opinion, you are exactly the same as your critic -> you are interested in making videos and selling your stuff since if you could you will make one 2 hour video covering all that you say properly, but... fakes don't do that, not lazy probably incompetent.
This is called a "strawman argument". Look it up. Nothing you've said is even remotely close to what I am saying. Stop inventing random things out of thin air.
@@LazyProgrammerOfficial Did you put attention to your name "Lazy Programmer". Your argument is called a "scammer argument".
@@abertj.7365 This video is about math, not trading. Do you know the difference? You can't even identify the *subject* of the video, never mind being able to intelligently discuss its contents. Stop embarrassing yourself.
@@LazyProgrammerOfficial So now you want to divert your laziness and say you preach about the correct math to one who holds Ph.D. in Math? Lazy one? Stop embarrassing yourself, you know nothing about math or trading, maybe you know some arithmetics and have seen some Hollywood movies about trading.
@@abertj.7365 You haven't even properly refuted any point in the video. Just throwing insults like a little man-child. Sad.
dislike this because promote course without evidence what's better
Sadly your comment is incomprehensible, can you try rephrasing? If you are sure of yourself, then please continue the discussion and tell me why I am wrong.
@@LazyProgrammerOfficial
This is not incomprehensible. You mentioned about another course isn't good at all, but after watch your video, you didn't show any improvement, or briliant approach to resolve the problem.
If you're kind of person who likes to be rewarded and praised then just delete my comment.
@@anhcoder Your thinking is incorrect (logical fallacy) and it's a problem which pervades research too. Showing that something is factually wrong doesn't require you to come up with a "better approach". A null result is a valid result. For instance, showing that stock prices closely follow a random walk indicates what further approaches should and should not be attempted.
> If you're kind of person who likes to be rewarded and praised
Sounds like you are taking the video as a personal attack (maybe you took this approach in the past?) instead of thinking scientifically.
@@anhcoder Can't come up with a rebuttal? ;)