Any thoughts on rolling z-scores or z-z-scores (z-scores of z-scores when distributions may be not as mean reverting) and testing for effective lookback periods without data mining and information creep?
You most definitley can do a rolling z-score of z scores approach. (In fact that approach can be an extension of what I did in this video.) In terms of back testing, that is definitley more of an artform that I am not 100% super familiar with. Many companies have their own built in backtesting. But, you can look at it from the point of view of specific time horizons in the stock exchange to judge how well your model will fare in economic upturns or dowturns. You don't have to have real time artificially generated data that represents a specific trend. At the end of the day, if you get good predictions, then by all means go for it! :)
Yes! I was essentially using the ratio between two different assets based on the historical data. And, based on the base model: rolling averages, I depend on the historical data to have a prediction in the future. (In essence, using in sample data for predictions) It is a basic model where the typical machine learning principles are not at play.
What is your rationale for doing RU-vid instead of developing a trading model? Do you view career or RU-vid as a more predictable way to increase earnings or do you doubt that it's possible to develop such model? Ignore question confines if needed, any thoughts appreciated.
An excellent question -- I actually never really put that much thought into creating quant models; I am more into Long only trades with longer time horizons. It may be an interesting venue to explore. It essentially boils down to time -- I only have so many hours in the day, and I recently became very, very busy. There's a lot more to that story but those are the high level points.
NOTE: at 4:20, pandas data reader has been deprecated!! So, the logic when obtaining your data will change. Thus, I have updated the Jupyter notebook here: github.com/SpencerPao/Pairs_Trading At 9:40 2 standard deviations are 95%. 1.25 standard deviations is ~90%
Yes, people do use returns for calculating hedging ratios in real life.. but I guess the reason he used "Price" is it is easier to create visualizations using Pricing data.
Hey Spencer, thankyou so much for the video!! This is my first video on pairstrading and I finally understood the idea. However I have a few doubts especially at 8:55 , when you calculate the mean and then find your price signals. When you do that does that not cause a look ahead bias for older signals? Like we really never knew it was the mean at the beginning of the datapoint right? I'm sorry for the dumb question, but if you can clarify it, that would help me a lot! Thankyou!
There are no dumb questions! :) There are inherent assumptions when you go and utilize the pair trading algorithm. (i.e what time horizon you want to choose) Bias is an inherent factor with most if not all time-series particular model. There are some mays to mitigate that bias: 1) you can have normalized models in relation to the market over differing economic trends (recession, depression, growth etc..) and see how that compares. 2) Have additional features that is more representative of the data you are seeing now 3) Data augmentation etc. Hopefully that answered your question?
Should note: you can have a cointegrated pair with zero correlation. Correlation does not imply cointegration or rather an underlying relationship which influences the process of each time series in a distinct way. The relationship is what we are trading, or in other more specific terms, the statistical deviation from a cointegrated spread.
Hi Spencer, this is a fantastic channel! Your explanations are some of the clearest I've seen. When you have the buying and selling signals, that's on the specific asset you declared you were going to be long or short one, correct? I would love a video on how to implement the actual buying/selling part of the strategy and backtesting if you're taking recommendations! Thanks again Spencer
To answer your first question -- Yes indeed! I'm glad that you enjoyed this video. :) I'll dabble in the back testing / actual implementation of the algorithm via a brokerage. and see if I can make a video quality walkthrough. Note that the back testing is an art form; many of the testing algorithms are in house. I'm sure that there are some open source backtesting algo's to use.
Sir, In pars trading can one trade spread between SP500 & Nasdq 100 based just on market sentiment of broad market/major indices.When sentiment is bullish,Nasdq100 would outperform SP 500 index.And in bearish scene vice versa.Same may be applied for pairs trading SP500 & Dow 30.In weak market Dow 30 would hold better & go down less compared to SP500/nasdq100.We seek your wisdom.Thank you.
Sir, In pair trading correlated currency pairs(say AUD/USD & NZD/USD) do you compare RSI value & sell overpriced pair & buy the weaker? OR do you compare CCI of each pair? Do you suggest using correlated cross currencies as pair trades(eg EUR/JPY & GBP/JPY)? On 4H chart sum up RSI 3, RSI 7,RSI 14, & compare added value on each pair & go. Thank you.
You can try any of those -- whichever has the best signals, go for it. Regardless of which relationships you seek, the pairs trading principle applies to any.
@@ronaknathwani5142 Hmm. Those are ticker symbols I assume? You can run any tool that you are comfortable with. (by tools do you mean software?) Check out code in C. Those are typically the fastest language in the quant world. If you mean which theories to use, you should test whichever one in your overall strategy to get your edge. Perhaps even use a combination of them or some other strategy that I am unfamiliar with.
@@ronaknathwani5142 Sir, On one chart plot RSI 3, RSI 7, RSI 14 ( say using MT4) & write values of each & add the numbers for ES NQ & YM--and compare.Please sell overpriced & buy the underpriced. System Programs use RSI 14, but as you know RSI 3 is the best guide to price & RSI 7 is more recent than RSI 14--so we have better picture of strength/weakness mtf.You need no program--just paper & pencil.
dude, you gotta get rid of this line in your code "Cointegration test: A technique used to find a potential correlation in a time series (long term)". This is *absolutely* not what it is. But you have mistaken correlation for causality the entire way through this video so I'm not surprised. Cointegration actually establishes a causal link from one time series to another" Peace bro. But please don't say the purpose of cointegration is to find correlations. Nothing is further from the truth :-)
Hi ! I tried to use your code however there is an error. From what I understand, it can't access the data from yahoo finance correctly. Are you experiencing the same issue ?
Yes indeed! Thank you for letting me know! Yahoo finance has changed their backend so the functions I go over in this video are deprecated. However, I have updated my notebook to adjust for this. You can check it out on my Github here: github.com/SpencerPao/Pairs_Trading
I tried to run it but sadly, the signal is not correct. As you can see on the graph u presented, the signal is false as well. Any other way to detect peak and bottom of a graph ?
Sir, Have you tried pair trading forex using rsi7 ,rsi14,rsi30 (add them up for comparison) say on hourly chart & selling strong pair & buying weak pair--pairs have to be highly correlated(eg aud/usd and nzd/usd OR dow30 & sp500).One can do this on any correlating underlying stocks/commodities/futures/crypto/bonds. Trading on hourly charts there would be tons of opportunities all year around.
I talk with others who are in the field. Honestly, the 'best' way to become a quant trader is to do it yourself. And, if you want to join a huge fund, you will have an edge over everyone else because you were an actual practioner. You learn by failing over and over.
Hey great video Spencer. I have a doubt at the end when we get the buy and sell signals, so say the buy signal is the trigger we got, now do we buy BRK or buy MSFT? how do we decide? Also how is it a pair strategy, we are only buying or selling only one stock at a time?
Mathematically and theoretically, you can go long on X and short Y. The direction of either one depends on what the model suggests. Also, from the pairs strategy (X,Y), you would typically standardize the pair i.e X would be buy and Y would be sell
I am curious about something, aren't we having a lookahead bias here when doing the tests to determine the pairs to trade, like say in one 1, we wont have info from the future right to tell us if this pair is cointegrated. How do you think we can tackle this?
Bias is huge. And, I don't talk about penalization. One way to do this is through continuous backtesting and revalidating the model (reweighting and reupdating the model) OR even having a different model altogether.
Spencer, very interesting video. Please note that there is an “overfitting” problem in your concept. You are using a prior knowledge regarding the correlation of these two instruments. In other words, the fact that these two stocks are correlated in the period of time that you are analyzing is not necessarily going to be true in the future (and you are analyzing the performance of your method only for the past signal… that was used to determine if the stocks are correlated in the first place). Take care
You're absolutley correct! This is an assumption that there is an inherent dependency in this method and your positions will inherently become more risky the longer you hold onto a specific position.
Hi Spencer thanks for the video. why did you choose 1.25 sd ? 1 sd from the mean covers 68% of the data I agree, but If you want to cover 95% of the data, you should choose 2 sd, considering a normal distribution. Thnaks
Most likely. If you are able to query data from the Binance APi without much limit, you can do the pairs trading algorithm. It's just a matter of latency at that point.
Goodnight ! Do you have any video on your channel about the Pair Trading strategy but instead of long x short, does it work in the form long & long or short & short? Note: congratulations for the channel.
Hello! I do not have a video on that topic. However, what you are stating is quite similar to the pair trading algorithm I discussed. Just switch the logic. I.e if asset A and asset B were to decrease (high correlation), you can spread your bets between the 2 asset classes.
It depends on your use case. If speed is not an issue, you can resort to scraping data from Yahoo Finance with Python/R. Those languages should suffice.
Ah LQE? Few folks are aware of that haha. Estimating position(s) is actually incredibly interesting based on uncertain features. That would be an entirely different video altogether (and not necessarily related to Pairs Trading) I have thought about using Kalman filter on an ARIMA model, and after reading a paper, the filter does provide an incremental improvement. I might do a video on the topic but not in the near future.
Algo system? Not really. For deployments for trading algorithms, you can look into using interactive brokers via their api. Plug in your model and stream their data. They are quite hefty in costs though.
@@amriksingh5003 Hmm. Well, unfortunatley everything has some elment of code. Though, you could just run my notebook locally from github and plug in your pair of public companies (for example) manually purchase/sell equity on some exchange.
@@ВиталийДягилев-ж5я ooh. I think you're referring to scraping financial asset data? I have not attempted to scrape from binance, so I am not familiar with the tooling, but you can set up price scraping mechanisms on Yahoo Finance's API. I did a similar video on this topic: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-GGgNM7WanK8.html Or this video (includes Yahoo finance data): ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-wlaLmM_LvWg.html
Hey, thanks for the amazing content, great video. I have one python-related technical question: by using pandas 1.2, the statsmodels doesn't work. I get this error when trying to import statsmodels.tsa.stattools: "module 'pandas' has no attribute 'Panel'". It seems the panel has been removed from Pandas module 0.25.0 onwards. How did you overcome this problem? Thanks in advance. Keep up the good work!