I've been successfully running a very similar model for the past 7 years . To improve the robustness of the list of stocks, a good filter to add would be to include only those stocks that not only have the best momentum, but also have greater momentum compared to the NASDAQ index. This will ensure that the stocks on the list not only have strong overall momentum, but also have momentum that is stronger than the broader market.
Great video, as always. A few points which I noticed after testing this system: - starting the backtest at 2005 hides a huge drawdon between 2000 and 2005 of > 60 % - The survivorship bias has huge impact on this system (as often for momentum systems): the CAGR decreases from 28 % to 15 % and the maxDD increases from 46 % to 65 % (tested from 2000 to today) - a simple regime filter helps a lot lowering the drawdowns - the system performs a little bit better on the S&P 500 universe with 250 stocks -> 100 stocks -> 10 stocks (not optimized)
What kind of regime filter do you use? And a more detailed question: as an example if you're using a 200sma, are you setting the filter based on the (in this case) Nasdaq 100 200sma and not trading when the overall market trend is down or the 200sma for each individual stock and picking the final 10 based on each being above it's indiviual 200sma?
I set the start date to 2000-01-01. Performance was pretty flat (just looking at the plot), but I don't see a huge drawdown? Perhaps it's too small to see when compared with more recent returns? I can see a 32% drawdown at 2009-01-31, but not one greater than 60%. Can you provide details?
Excellent video as always. I made some mods using the Russel 1000, and instead of equal weights, I run PypfOpt Efficient Frontier and Discrete Allocation in every selection, optimizing for Sharpe ratio and using exponentially weighted covariance matrix. For 2022 alone, cumprod goes from 1.10 to 1.28 doing that. Even more interesting, I ran it weekly (26, 13 and 6 weeks), which gives me a cumprod of 1.30 with mean, and 1.57 with discrete allocation. I'll keep experimenting. Thanks for the great quality work you put out!
I did this in an Excel spreadsheet starting in 1993. It was based on a mutual fund portfolio in my 401(k) account. I had excellent returns during those years. The strategy was to move money toward funds that were increasing in momentum, While pulling money out of funds that were decreasing in momentum. If most funds were decreasing, the strategy was to park the money into stable value fundS. Of course the brokerage rules changed as time went on, which meant mandatory longer wait periods of time between trades. Evolved Very much like your monthly analysis period example. My original strategy, was one month 15 days five days.
If there is anything that is missing on the internet, it's a real strategy that uses the ichimoku cloud in the right way! I tried to get chat gpt to make it but it had no idea how to create it. The real icumoku cloud has 2 different ways to give you a signal. First one is : Price above the cloud, lagging span above the price, and conversion line crossing the baseline above the cloud, then a green portion of the cloud 26 periods in front of the price.. That is a trend continuation signal. Then you have the pump signal: Price above the cloud, green portion of the cloud 26 period in front, cross below or inside the cloud, lagging span above the cloud
I have made half of that in pine script if you want to have a look at it, but it's not that good... I'm quite shitty as a programmer. But i have some ideas. 😉
This example clearly shows how fees killing the most strategy at 22:22. Even with just 1 percent of fee, it went from 20x to 5x… ehh. I liked the content though, great video!
Hi mate, thanks a lot for watching! Agreed on the one hand, on the other hand 1% is way too much. I took a way to conservative approach here, that's why it is crushing the PnL to that extent.
Great work, thanks for your video AlgoVibes, I'm waiting for the fully SVS-bias free version, and, if possible, with the eur/usd conversion (selectable as a choice). Regarding the drawdown problem, I recommend the book in which Andreas Cleanow presents his approach :"Stocks on the move". Here the momentum strategy presented (on SP500) is a bit more complex. But a simple SMA 200 filter on the benchmark is sufficient to halve the DD, with a very limited reduction in the CAGR.
I liked your code very much! So clean. The performance of the strategy is almost the same as the index. Strategy performance - Market performance < Headache of maintaining a trading bot
Thanks mate. Yeah but you have to consider that I filtered out the astronomic performers like Air BNB and TSLA to diminish the survivorship bias. That will ofc make the performance way worse then it actually would have been.
Excellent video. Very comprehensive. Appreciate your patience in the presentation. A few comments: 1. I am using Jupyter Notebook. I was unable to use pd.read_html to access the list of the 100 NASDAQ stocks. Ending up copying the table to an Excel spreadsheet, save in .csv format then load the data using pd.read_csv. 2. Interesting to see what the result of a momentum based long-short strategy. 3. Instead of comparing strategy results to IXIC, I suggest comparing results to an equal weighted portfolio of the surviving stocks.
thanks for a great presentation. Kindly consider adding code to mitigate the survivorship bias issue. Also adding limiting drawdown % as suggested in the comments below, myself i was using a 200SMA filter. I am currently working on a momentum strategy but had difficulty in determining what time windows to use so this video was invaluable to me. Thanks to Federico as well.
Nice strategy and presentation! As an easy approach to consider the survivorship bias, could it be that you may just replace NaN with 1.0 (like holding cash). This would also mean that in bad months where lots of shares have returns < 1, you start holding cash and only hold shares with return > 1.
Thanks mate. The problem with my approach is, that I fully exclude them from my analysis as I do not know when the came into the index. There is some point in time when they climbed in the index and from this point in time on I have to consider them as well.
Hey @AlgoVibes, loved the video. What a beautifully written and well explained code. Just one question. What date would you pass each month in the last Get_Top(Date) function to get names of 10stocks to invest in? Would it be starting range of the data or the current date where you want recommendations per the model? I couldn’t understand that part of the logic. TIA
well done ! great job! for the benchmarks and performance a lot of the time that is expressed with the sharpe ratio as you can later leverage the strategy as you wish
@@Algovibes test for low volatility anomaly based on past 36 months returns standard deviation (or beta), and rebalance on monthly or quarterly basis. I tested for S&P 500 and getting opposite results from what academic research suggests. So want to see how others are implementing the strategy 🙏
Hi Algovibes, i would be glad about an implementation of an regime filter e.g. QQQ has to be over 200 SMA or 3 Month-Perf of QQQ has to be over 0 %. About this Video: For me (a real beginner) was it much easier to follow this video than with the last momentum update video with the S&P500. Without to understand everything, this time the programming seems like more straight forward (man brech ich mir auf englisch einen ab ;) Thank you
06:43 i had to add df.index = pd.to_datetime(df.index) in the line before mtl = (df.pct_change() +1)[1:].resample('M').prod() otherwise resample didn't work
In order to remove the survivorship bias, where would you get a dataset of the constituents of NASDAQ over time? It would be great if you could make a video about that. Great work, super helpful!
Yep! Chris is 100% right (unfortunately). When doing academic research you are usually working with the CRSP database which is not available for free. My idea was to take a table showing the changes in the last years and merge that table with the strategy. Table doesn't go back too much into the past but this is at least taking out the survivorship bias for the last decade.
Great video presentation! Enjoyed both the strategy and the Python implementation. Would you consider doing a follow on video where you vary the holding period of the top 10 stocks? Say for 3 months or 6 months before you refresh with a new set of stocks? Just curious whether backtesting such a strategy would improve the performance. Also interested to see how this could be implemented in Python. Thanks! (Enjoyed subscribing to your channel and learning from the other topics you've covered with Python!)
Hi have you ever thought about trying a strategy that takes volatility into account? Adapting position sizing or being/not being in a position based on how much volatility is on that particular asset? Also similar considerations can be made on volumes.... a video on that would be great! have a good day!
Hey Luigi, yes there are actually older videos on my channel that but in my newest video I am using the Sharpe ratio as my optimization objective. Stay tuned! :-)
Normalize position sizing accounting for volatility. And scale out if the trade goes against you, ie: sell a third of the position if lose x% or monthly low
Thanks mate. Well that would be actually quite an interesting idea. Kind of find the optimal lookback/holding period out of the cascade. Thanks for your input!
I don't seem to manage to upload the libraries such as pandas or finance in Jupiter Notebook. I have to use PyCharm that works. How do you get these libraries t work in Jupiter? I've tried launching Notebook via Anaconda, but I have the same issue... thx
Thanks mate! You can surely backtest momentum on commodities. The index component side (pick best of index) of this strategy I see quite difficult to mirror with commodities.
Thanks for the video! I found it interesting that if you held an equal weight portfolio from the inception of the strategy you outperform the cap weighted index by a factor of 2x `(df.dropna(axis=1).sum(axis=1).pct_change() + 1).cumprod().plot()` . Am I thinking about this properly?
which version of your python is ? i've tried to type and install yfinance in python 3.11, but it failed to comply and kept telling me there is syntax error.
Hi Dale, love your coding style and share your passion for algo trading but in the end the algos which incorporate chart pattern recognition will rule the future.
I was going to ask you why you don't subtract one from MTL but it looks like you want to maintain (1+r) so that when you apply the rolling function we can, for example take the last n=window=12 months within the get_rolling_ret(df,n) function. feeding to the function seems to provide you (1+r)_1 * (1+r)_2 * ...(1+r)_12 correct? Also, ret_12, ret_6, and ret_3 could have been calculated in a triple loop but this way is much more efficient correct?
@@Algovibes I did the portfolio sort with GICS industries AND the same idea. I may turn this into a class but I'm going to add a conditional via volatiility and observe results. I also think trying to combine momentum with a short-term reversal would be interesting. FYI The optimization you're about to do is going to interesting to see from a looping perspective. If you only use the quadratic and linear constraint (return and variance with no other constratins) it's not going to make any sense. You may want to add a max trade size condition.
Thanks for this. Could anyone help me to understand why removing the stocks that were not on NASDAQ from the chosen data would help to overcome survivorship bias?
Welcome man! I just released a new video explaining and coding survivorship bias. That should answer your question hopefully. If not - pls let me know.
Thank you for this ! Can you tell me how can I apply the risk managed momentum formula from the paper Momentum has its moments (Barroso, Santa-Clara) to this ? Moreover, for comparison purposes I would like to combine the cross sectional and time series momentum to see which one performs better… do you have any idea ? Thanks in advance
Even better: You can become a channel member, support the channel with that and get access to the code and discord :-) Be invited to join here: th-cam.com/channels/87aeHqMrlR6ED0w2SVi5nw.htmljoin
im sort of stuck on something. I was testing this out with a different set of stocks. The portoflio returns im getting, at the end, are all the same. The return is the same on every day.
Thanks buddy, agreed but in the end I am just setting an index for the series. Nothing complicated. But I was making it way more complicated then it actually is.
On the other hand, this strategy would certainly seem likely to ride a stock like TSLA. Just saying, I wouldn't count another outlier like that. Would be interesting to see how this strategy performs with the S&P500 before TSLA was added to it.
Hi algovibes, there is a github by a gentleman Teddy Koker that discusses how to construct a sp500 dataset free of survivorship bias, using free data sources. Perhaps you can cover this in your next video ?
Hi buddy, thanks a lot for the suggestion. Anyhow if you check out his note: "EDIT 2020: iShares no longer publishes historical holdings to their website (perhaps because of this abuse). Constituents until mid-2019 are available in constituents.csv"
@@Algovibes hi, i suppose it is still possible to update the 2019 - 2022 constituents by hand since the sp500 Wikipedia would have captured the changes. Even with a dataset dating to 2019 , it would be helpful if you can do a video to guide us on backtesting a momentum strategy on a dataset free of survivorship bias.
Hi Algovibes, great video as usual but it still has survivorship bias as some of the companies might not be in Nasdaq 100 10 years back, they might have replaced some companies and made it in to list. Another point when I start it from 2000, I might again remove some of the great companies of that time.
Is the strategy idea only to look at the last year of data? For instance if I wanted to pick 10 stocks for my portfolio for this month, should I only include 2022 data or does this historical going back to 2010 (or beyond) matter for accuracy?
this is a momentum strategy. It looks at 3 time frames 12 month, 6 month, and 3 month. And chooses 10 stocks to buy and hold for a month. Its back tested on data going back to 2010 which gives you those returns.
Hi, for everything related to return calculation please watch this video: th-cam.com/video/fWHQwqT3lNY/w-d-xo.html Let me know if there is anything unclear left!
Great video, but it is not taking in consideration that in many countries for each winning position that is closed you have to give a cut to the government as taxes
Great video! I am trying to replicate, but I get performance value for first months also, not NaN as it should. At what point do you make the check for if there are returns for 12, 6, 3 months? With .index you get stock names and from mtl I get " Date 2010-02-28 1.040044" for the 10 stocks from index. I shouldn't get that while 12,6 and 3 months don't have any returns. I have it like this: def pf_performance(date): portfolio = mtl.loc[date:, get_top(date)][1:2] return portfolio.mean(axis=1).values[0] returns = [] for date in mtl.index[:-1]: returns.append(pf_performance(date))
Whats the practical use of that so-called strategy? To check how much could you earn if invested in those stocks 12 years ago? Very informative as a coding example but useless as a trading strategy)
Hi Johannes, you need to grab a coffee ☕️, sit back and watch the whole video again. I clearly stated in the video I left out the survivorship bias part as well as clearly communicated the impact of dropping the NaNs. Don't worry about the strategy guy. Federico already thanked me and was 100% satisfied. By the way: Leave a like and share the video for the survivorship bias part!
You fools who praise this video. First, it has a wrong calculation and, secondly, it doesn't take a decreasing nasdaq into account like it has been in 2009. The calculation starts at 2010. But anyhow, if you overcome these issues the video is helpful. If you get the calculation right you get about -12% in the last year. If you improve more then you can get +12% out of each year with 0.01 substracted from profits.
I strongly disagree on this comment just for the record. Nobody is a fool for praising high quality content from an actual expert in the field. Calculations are correct and lookback is more than enough. Survivorship bias is an issue, which has been covered here: th-cam.com/video/OHYExP6RVFU/w-d-xo.html
#1 mistake when testing on historical stock data - survivor ship bias, which would drastically affect the performance of a strategy like this; the real-world performance would be around 30% worse than presented in this video
Sorry to hear but I won't write a script. I played around with scripting but people clearly prefer it freestyle as it seems. Besides that I don't have time for setting up a script - it simply isn't worth it. Thanks for your understanding!
I tried to step by step but I got vaild error.I need some help ! Would you mind solve the problem? What's wrong? Thank you so much. File ~\anaconda3\lib\site-packages\spyder_kernels\py3compat.py:356 in compat_exec exec(code, globals, locals) File c:\users\py asdaq momentum.py:50 portfolio = mtl.loc['2010-12-31':,get_top][1:2] File ~\AppData\Roaming\Python\Python39\site-packages\pandas\core\indexing.py:1094 in __getitem__ key = tuple(com.apply_if_callable(x, self.obj) for x in key) File ~\AppData\Roaming\Python\Python39\site-packages\pandas\core\indexing.py:1094 in key = tuple(com.apply_if_callable(x, self.obj) for x in key) File ~\AppData\Roaming\Python\Python39\site-packages\pandas\core\common.py:379 in apply_if_callable return maybe_callable(obj, **kwargs) File c:\users\py asdaq momentum.py:41 in get_top top_50 = ret_12.loc[date].nlargest(50).index File ~\AppData\Roaming\Python\Python39\site-packages\pandas\core\indexing.py:1103 in __getitem__ return self._getitem_axis(maybe_callable, axis=axis) File ~\AppData\Roaming\Python\Python39\site-packages\pandas\core\indexing.py:1330 in _getitem_axis raise ValueError("Cannot index with multidimensional key") ValueError: Cannot index with multidimensional key import yfinance as yf import pandas as pd import matplotlib.pyplot as plt import numpy as np ticker_df = pd.read_html("en.wikipedia.org/wiki/Nasdaq-100")[4] tickers = ticker_df.Ticker.to_list() df = yf.download(tickers,start='2010-01-01')['Adj Close'] print(df) df = df.dropna(axis=1) mtl = (df.pct_change() +1)[1:].resample('M').prod() print(mtl) def get_rolling_ret(df,n): return df.rolling(n).apply(np.prod) ret_12, ret_6, ret_3 = get_rolling_ret(mtl,12),get_rolling_ret(mtl,6) ,get_rolling_ret(mtl,3) print(ret_12) print(ret_6) print(ret_3) top_50 = ret_12.loc['2010-12-31'].nlargest(50).index top_30 = ret_6.loc['2010-12-31', top_50].nlargest(30).index top_10 = ret_3.loc['2010-12-31', top_30].nlargest(10).index print(top_50) print(top_30) print(top_10) def get_top(date): top_50 = ret_12.loc[date].nlargest(50).index top_30 = ret_6.loc[date, top_50].nlargest(30).index top_10 = ret_3.loc[date, top_30].nlargest(10).index return top_10 get_top('2010-12-31') mtl.loc['2010-12-31':][1:2] portfolio = mtl.loc['2010-12-31':,get_top][1:2] portfolio.mean(axis=1).values[0] def pf_performance(date): portfolio = mtl.loc[date:, get_top(date)][1:2] return portfolio.mean(axis=1).values[0] pf_performance('2010-12-31') mtl.index[:-1] print(mtl) returns = [] for date in mtl.index[:-1]: returns.append(pf_performance(date)) mtl.index[1:] pd.Series([i-0.01 for i in returns],index=mtl.index[1:]).cumprod().plot() pd.Series([i-0.01 for i in returns],index=mtl.index[1:]).cumprod() nas_df = yf.download('IXIC', '2010-01-01') print(nas_df) (nas_df['Adj Close'].pct_change() +1).cumprod() (nas_df['Adj Close'].pct_change() +1).cumprod().plot() pd.Series([i-0.01 for i in returns],index=mtl.index[1:]).cumprod().plot() portfolio = mtl.loc['2010-12-31':,get_top][1:2] portfolio.mean(axis=1).values[0] def pf_performance(date): portfolio = mtl.loc[date:, get_top(date)][1:2] return portfolio.mean(axis=1) pf_performance('2010-12-31') mtl.index[:-1] print(mtl) returns = [] for date in mtl.index[:-1]: returns.append(pf_performance(date)) mtl.index[1:] pd.Series([i-0.01 for i in returns],index=mtl.index[1:]).cumprod().plot() pd.Series([i-0.01 for i in returns],index=mtl.index[1:]).cumprod() nas_df = yf.download('IXIC', '2010-01-01') print(nas_df) (nas_df['Adj Close'].pct_change() +1).cumprod() (nas_df['Adj Close'].pct_change() +1).cumprod().plot() pd.Series([i-0.01 for i in returns],index=mtl.index[1:]).cumprod().plot()
![How to build a trading strategy [Momentum] with Python?](http://i.ytimg.com/vi/dnrJ4zwCADM/mqdefault.jpg)
I've been successfully running a very similar model for the past 7 years .
To improve the robustness of the list of stocks, a good filter to add would be to include only those stocks that not only have the best momentum, but also have greater momentum compared to the NASDAQ index. This will ensure that the stocks on the list not only have strong overall momentum, but also have momentum that is stronger than the broader market.
Cool! Thanks a lot for sharing man.
What is your return?
.
i dint understand. how could a stock have the best momentum and at the same time have a lower momentum than the index?
Check that the momentum is greater than 0 could be a good addition to stop drawdowns.
Great video, as always. A few points which I noticed after testing this system:
- starting the backtest at 2005 hides a huge drawdon between 2000 and 2005 of > 60 %
- The survivorship bias has huge impact on this system (as often for momentum systems): the CAGR decreases from 28 % to 15 % and the maxDD increases from 46 % to 65 % (tested from 2000 to today)
- a simple regime filter helps a lot lowering the drawdowns
- the system performs a little bit better on the S&P 500 universe with 250 stocks -> 100 stocks -> 10 stocks (not optimized)
Good comment. Thanks a lot Andy!
What kind of regime filter do you use?
And a more detailed question: as an example if you're using a 200sma, are you setting the filter based on the (in this case) Nasdaq 100 200sma and not trading when the overall market trend is down or the 200sma for each individual stock and picking the final 10 based on each being above it's indiviual 200sma?
I set the start date to 2000-01-01. Performance was pretty flat (just looking at the plot), but I don't see a huge drawdown? Perhaps it's too small to see when compared with more recent returns? I can see a 32% drawdown at 2009-01-31, but not one greater than 60%. Can you provide details?
Yes would like to see how it changes without survivorship bias.
Thanks a lot for your feedback! Will see what I can do.
Excellent video as always. I made some mods using the Russel 1000, and instead of equal weights, I run PypfOpt Efficient Frontier and Discrete Allocation in every selection, optimizing for Sharpe ratio and using exponentially weighted covariance matrix. For 2022 alone, cumprod goes from 1.10 to 1.28 doing that. Even more interesting, I ran it weekly (26, 13 and 6 weeks), which gives me a cumprod of 1.30 with mean, and 1.57 with discrete allocation. I'll keep experimenting. Thanks for the great quality work you put out!
Thank you very mate for both the appreciative words and also sharing your strategy!
I did this in an Excel spreadsheet starting in 1993. It was based on a mutual fund portfolio in my 401(k) account. I had excellent returns during those years. The strategy was to move money toward funds that were increasing in momentum, While pulling money out of funds that were decreasing in momentum. If most funds were decreasing, the strategy was to park the money into stable value fundS. Of course the brokerage rules changed as time went on, which meant mandatory longer wait periods of time between trades. Evolved Very much like your monthly analysis period example. My original strategy, was one month 15 days five days.
Cool man, thanks a ton for sharing. How did it work out? 😄
@@Algovibes I retired at 59.
hello Mr david, how can you speak with you ? (gmail,whatsapp number or something like that)
Definitely be interested in seeing this without any survivorship bias
Already did it! Be invited to check it out :-)
If there is anything that is missing on the internet, it's a real strategy that uses the ichimoku cloud in the right way! I tried to get chat gpt to make it but it had no idea how to create it.
The real icumoku cloud has 2 different ways to give you a signal.
First one is : Price above the cloud, lagging span above the price, and conversion line crossing the baseline above the cloud, then a green portion of the cloud 26 periods in front of the price.. That is a trend continuation signal.
Then you have the pump signal:
Price above the cloud, green portion of the cloud 26 period in front, cross below or inside the cloud, lagging span above the cloud
I have made half of that in pine script if you want to have a look at it, but it's not that good... I'm quite shitty as a programmer. But i have some ideas. 😉
Hey Mikul, really appreciate your comment. Sorry for my late reply. I noted that for potential future content.
This example clearly shows how fees killing the most strategy at 22:22. Even with just 1 percent of fee, it went from 20x to 5x… ehh. I liked the content though, great video!
Hi mate, thanks a lot for watching!
Agreed on the one hand, on the other hand 1% is way too much. I took a way to conservative approach here, that's why it is crushing the PnL to that extent.
Great work, thanks for your video AlgoVibes, I'm waiting for the fully SVS-bias free version, and, if possible, with the eur/usd conversion (selectable as a choice). Regarding the drawdown problem, I recommend the book in which Andreas Cleanow presents his approach :"Stocks on the move". Here the momentum strategy presented (on SP500) is a bit more complex. But a simple SMA 200 filter on the benchmark is sufficient to halve the DD, with a very limited reduction in the CAGR.
Wow man! Good stuff! Thank you!
Thanks buddy :-)
I liked your code very much! So clean.
The performance of the strategy is almost the same as the index.
Strategy performance - Market performance < Headache of maintaining a trading bot
Thanks mate. Yeah but you have to consider that I filtered out the astronomic performers like Air BNB and TSLA to diminish the survivorship bias. That will ofc make the performance way worse then it actually would have been.
Awesome buddy. The wait was worth.
thanks mate :-)
Excellent video. Very comprehensive. Appreciate your patience in the presentation. A few comments:
1. I am using Jupyter Notebook. I was unable to use pd.read_html to access the list of the 100 NASDAQ stocks. Ending up copying the table to an Excel spreadsheet, save in .csv format then load the data using pd.read_csv.
2. Interesting to see what the result of a momentum based long-short strategy.
3. Instead of comparing strategy results to IXIC, I suggest comparing results to an equal weighted portfolio of the surviving stocks.
Hi buddy, very good points. Thanks a lot for sharing!
thanks for a great presentation. Kindly consider adding code to mitigate the survivorship bias issue. Also adding limiting drawdown % as suggested in the comments below, myself i was using a 200SMA filter. I am currently working on a momentum strategy but had difficulty in determining what time windows to use so this video was invaluable to me. Thanks to Federico as well.
Welcome mate! Happy you could extract value out of the video.
I always look forward to your videos, thanks
Thanks for your support Pavel! Really appreciate it.
Hi! My favourite videos style, old factors tested! Glad to see this return to the spots. Thanks
Hi 🙋🏻 Awesome. Love to make these videos as well. Thanks for watching rraul.
Thank you. Yes, without survivorship bias would be great.
@@rame2rame Testing strategies with survivorship bias delivers unrealistic results, because we only test on securities that have survived.
@@ytuser7967 thanks for your feedback mate
@@Algovibes Your are very welcome.
Nice strategy and presentation! As an easy approach to consider the survivorship bias, could it be that you may just replace NaN with 1.0 (like holding cash). This would also mean that in bad months where lots of shares have returns < 1, you start holding cash and only hold shares with return > 1.
Thanks mate. The problem with my approach is, that I fully exclude them from my analysis as I do not know when the came into the index. There is some point in time when they climbed in the index and from this point in time on I have to consider them as well.
@@Algovibes So? Just replace NaN with 1.
Hey @AlgoVibes, loved the video. What a beautifully written and well explained code. Just one question. What date would you pass each month in the last Get_Top(Date) function to get names of 10stocks to invest in? Would it be starting range of the data or the current date where you want recommendations per the model? I couldn’t understand that part of the logic. TIA
Great stuff, small suggestion here - would be great to plot 2 equity curves on one graph, also showing the return% on y and date on x.
Thanks buddy. Nice suggestion!
I was wondering if you can create a video on dual-momentum strategy as per Gary Antonacci's book.
well done ! great job! for the benchmarks and performance a lot of the time that is expressed with the sharpe ratio as you can later leverage the strategy as you wish
Thanks a ton Ricardo. Yep! Check out the newest video. I am talking about Sharpe ratio there. Sharpe ratio is usually the optimizing objective.
Hi Algovibe. Can you also test low volatility strategy on NASDAQ (or any stock universe)? Curious to see the strategy results
Interesting suggestion! Anything specific you have in mind?
@@Algovibes test for low volatility anomaly based on past 36 months returns standard deviation (or beta), and rebalance on monthly or quarterly basis. I tested for S&P 500 and getting opposite results from what academic research suggests. So want to see how others are implementing the strategy 🙏
the full survivorship bias video would be great!!
Thanks for your feedback mate!
Hi Algovibes,
i would be glad about an implementation of an regime filter e.g. QQQ has to be over 200 SMA or 3 Month-Perf of QQQ has to be over 0 %. About this Video: For me (a real beginner) was it much easier to follow this video than with the last momentum update video with the S&P500. Without to understand everything, this time the programming seems like more straight forward (man brech ich mir auf englisch einen ab ;)
Thank you
Thanks for your feedback!
Und außerdem Danke für deine sprachlichen Bemühungen! :-)
maybe do the program on bigger unverse of stocks out of the index would minimimse survivalship bias
It is really cool. Python makes the coding simple.
Yep :-)
06:43 i had to add df.index = pd.to_datetime(df.index) in the line before mtl = (df.pct_change() +1)[1:].resample('M').prod() otherwise resample didn't work
In order to remove the survivorship bias, where would you get a dataset of the constituents of NASDAQ over time? It would be great if you could make a video about that. Great work, super helpful!
This data is not free unless you create it yourself. Norgate is a data provider that can provide this to you.
Yep! Chris is 100% right (unfortunately). When doing academic research you are usually working with the CRSP database which is not available for free. My idea was to take a table showing the changes in the last years and merge that table with the strategy. Table doesn't go back too much into the past but this is at least taking out the survivorship bias for the last decade.
Great video presentation! Enjoyed both the strategy and the Python implementation. Would you consider doing a follow on video where you vary the holding period of the top 10 stocks? Say for 3 months or 6 months before you refresh with a new set of stocks? Just curious whether backtesting such a strategy would improve the performance. Also interested to see how this could be implemented in Python. Thanks! (Enjoyed subscribing to your channel and learning from the other topics you've covered with Python!)
Thanks a lot Jonny! Cool idea also.
You are an angel fallen from the heaven. Thank you
Same is true for you leaving a comment like that :-) Thanks for your support man.
Hi have you ever thought about trying a strategy that takes volatility into account? Adapting position sizing or being/not being in a position based on how much volatility is on that particular asset? Also similar considerations can be made on volumes.... a video on that would be great! have a good day!
Hey Luigi,
yes there are actually older videos on my channel that but in my newest video I am using the Sharpe ratio as my optimization objective. Stay tuned! :-)
Normalize position sizing accounting for volatility. And scale out if the trade goes against you, ie: sell a third of the position if lose x% or monthly low
Great Video again , Sir Waiting for your New Video
Thanks my man! New video is in production.
Great video. I would love to see the next step in optimization. Keep it up!
Thanks mate. Well that would be actually quite an interesting idea. Kind of find the optimal lookback/holding period out of the cascade. Thanks for your input!
survivorship bias full excluder would be great!!
Thanks for your feedback!
Keep us amazed buddy, and thank you so much for these ideas. I have a strategy which i would want to test it. How is it possible to reach you?
Thanks mate :-) Sure, just drop me a mail (can be find in about me) and I'll see what I can do!
I really like your videos. Thank you so much. They help me
Thank you very much Monique!
I don't seem to manage to upload the libraries such as pandas or finance in Jupiter Notebook. I have to use PyCharm that works. How do you get these libraries t work in Jupiter? I've tried launching Notebook via Anaconda, but I have the same issue... thx
Hello...this video was awesome! I have no experience with coding at all but i assume this could be done with commodities and futures correct?
Thanks mate! You can surely backtest momentum on commodities. The index component side (pick best of index) of this strategy I see quite difficult to mirror with commodities.
Thanks for the video! I found it interesting that if you held an equal weight portfolio from the inception of the strategy you outperform the cap weighted index by a factor of 2x `(df.dropna(axis=1).sum(axis=1).pct_change() + 1).cumprod().plot()` . Am I thinking about this properly?
Another superb video, with lots of information and knowledge shared along with python coding learning. Thanks dear 🙏🙏
Thanks for watching mate!
Great video. I was wondering does the strategy assumes we are holding the stock or when do we exit?
Thanks mate! Was taking a holding period of 1 month here.
Will this strategy perform in the S&P 500 or Nikkei 225?
Good question. Has to be tested. You have everything what you need to test it so go for it 😛
which version of your python is ? i've tried to type and install yfinance in python 3.11, but it failed to comply and kept telling me there is syntax error.
Hi Dale, love your coding style and share your passion for algo trading but in the end the algos which incorporate chart pattern recognition will rule the future.
You make great videos.Could you make a video with a strategy on renko chart!?
Thanks a ton for your suggestion and your comment mate :-)
I was going to ask you why you don't subtract one from MTL but it looks like you want to maintain (1+r) so that when you apply the rolling function we can, for example take the last n=window=12 months within the get_rolling_ret(df,n) function. feeding to the function seems to provide you (1+r)_1 * (1+r)_2 * ...(1+r)_12 correct?
Also, ret_12, ret_6, and ret_3 could have been calculated in a triple loop but this way is much more efficient correct?
Hi Bryan, Yep! You got it 100% right. Nothing to add :-)
@@Algovibes I did the portfolio sort with GICS industries AND the same idea. I may turn this into a class but I'm going to add a conditional via volatiility and observe results. I also think trying to combine momentum with a short-term reversal would be interesting. FYI The optimization you're about to do is going to interesting to see from a looping perspective. If you only use the quadratic and linear constraint (return and variance with no other constratins) it's not going to make any sense. You may want to add a max trade size condition.
Thanks for this.
Could anyone help me to understand why removing the stocks that were not on NASDAQ from the chosen data would help to overcome survivorship bias?
Welcome man! I just released a new video explaining and coding survivorship bias. That should answer your question hopefully. If not - pls let me know.
@@Algovibes great! Thanks! Will watch and if I have questions I’ll post them as comments.
Thank you for this !
Can you tell me how can I apply the risk managed momentum formula from the paper Momentum has its moments (Barroso, Santa-Clara) to this ?
Moreover, for comparison purposes I would like to combine the cross sectional and time series momentum to see which one performs better… do you have any idea ?
Thanks in advance
Nice lecture! Are you able to post the python code? Thanks.
Even better: You can become a channel member, support the channel with that and get access to the code and discord :-)
Be invited to join here:
th-cam.com/channels/87aeHqMrlR6ED0w2SVi5nw.htmljoin
im sort of stuck on something. I was testing this out with a different set of stocks. The portoflio returns im getting, at the end, are all the same. The return is the same on every day.
Hey Dennis,
If you elaborate a bit more on your problem I could help you. Or just post the problem in the Discord. People may solve it for you!
Shouldn't you be working with log returns? backtest not accurately reflecting performance with percentage changes, right?
Quiet straightforward answer: No.
When and why you should use log returns is explained here:
th-cam.com/video/fWHQwqT3lNY/w-d-xo.html
very good video, but the dates part at the end is a bit confusing
Thanks buddy, agreed but in the end I am just setting an index for the series. Nothing complicated. But I was making it way more complicated then it actually is.
TSLA probably causing major skew. Would be interesting to run excluding TSLA outlier.
Edit: Oh I see you address this at the end
On the other hand, this strategy would certainly seem likely to ride a stock like TSLA. Just saying, I wouldn't count another outlier like that. Would be interesting to see how this strategy performs with the S&P500 before TSLA was added to it.
Hi algovibes, there is a github by a gentleman Teddy Koker that discusses how to construct a sp500 dataset free of survivorship bias, using free data sources. Perhaps you can cover this in your next video ?
Hi buddy, thanks a lot for the suggestion. Anyhow if you check out his note:
"EDIT 2020: iShares no longer publishes historical holdings to their website (perhaps because of this abuse). Constituents until mid-2019 are available in constituents.csv"
@@Algovibes hi, i suppose it is still possible to update the 2019 - 2022 constituents by hand since the sp500 Wikipedia would have captured the changes.
Even with a dataset dating to 2019 , it would be helpful if you can do a video to guide us on backtesting a momentum strategy on a dataset free of survivorship bias.
Hi Algovibes, great video as usual but it still has survivorship bias as some of the companies might not be in Nasdaq 100 10 years back, they might have replaced some companies and made it in to list. Another point when I start it from 2000, I might again remove some of the great companies of that time.
Thanks mate. Yes, that's right. If you watch the video again I stated that in the video 😁
Is the strategy idea only to look at the last year of data? For instance if I wanted to pick 10 stocks for my portfolio for this month, should I only include 2022 data or does this historical going back to 2010 (or beyond) matter for accuracy?
this is a momentum strategy. It looks at 3 time frames 12 month, 6 month, and 3 month. And chooses 10 stocks to buy and hold for a month. Its back tested on data going back to 2010 which gives you those returns.
Thanks @PatzyBeats that's right!
Please show the version without Surivorship Bias
Sure! Thanks for your feedback :-)
Great video
Thanks mate!
What Python Framework are you using, it seems Google Collaborate, but looks faster!
Just Jupyter Notebook :-) Thanks for watching mate
Why did you add 1 with pct change ?
Hi, for everything related to return calculation please watch this video:
th-cam.com/video/fWHQwqT3lNY/w-d-xo.html
Let me know if there is anything unclear left!
Hi. How I can contact you? we need an educational lesson
Hi! I am currently quite swamped but you are invited to drop me a mail. You will find this in the About section on my channel.
@@Algovibes I wrote you on 27 november
Great video, but it is not taking in consideration that in many countries for each winning position that is closed you have to give a cut to the government as taxes
Super cool content
Happy to read. Thanks a lot :-)
Which python IDE is he using ?
Using Jupyter Notebook here.
Great video! I am trying to replicate, but I get performance value for first months also, not NaN as it should. At what point do you make the check for if there are returns for 12, 6, 3 months?
With .index you get stock names and from mtl I get " Date
2010-02-28 1.040044" for the 10 stocks from index. I shouldn't get that while 12,6 and 3 months don't have any returns.
I have it like this:
def pf_performance(date):
portfolio = mtl.loc[date:, get_top(date)][1:2]
return portfolio.mean(axis=1).values[0]
returns = []
for date in mtl.index[:-1]:
returns.append(pf_performance(date))
Is it safe to combine stocks from multiple indexes ?
Can you elaborate? thx!
@@Algovibes Is it safe to mix stocks from several indexes and build a lookback portfolio from all these mixed stocks?
If a security reaches monthly new highs buy a little. Then if it reaches three month new highs buy a little more.
Cool idea!
pip install pandas==1.3.0.................for those who cannot run resample for bugs in other pandas
I'm blown by the Python. What you can do with lots of coding is being done in few lines. Must learn for quick analysis.
Yes it's awesome. Thanks for your comment my man :-)
Whats the practical use of that so-called strategy? To check how much could you earn if invested in those stocks 12 years ago? Very informative as a coding example but useless as a trading strategy)
The strategy guy’s goal was clearly not met. Excluding NAs has no effect at all and the backtest has 100% survivorship and look-ahead bias in it.
Hi Johannes, you need to grab a coffee ☕️, sit back and watch the whole video again.
I clearly stated in the video I left out the survivorship bias part as well as clearly communicated the impact of dropping the NaNs.
Don't worry about the strategy guy. Federico already thanked me and was 100% satisfied.
By the way: Leave a like and share the video for the survivorship bias part!
You fools who praise this video. First, it has a wrong calculation and, secondly, it doesn't take a decreasing nasdaq into account like it has been in 2009. The calculation starts at 2010. But anyhow, if you overcome these issues the video is helpful. If you get the calculation right you get about -12% in the last year. If you improve more then you can get +12% out of each year with 0.01 substracted from profits.
boo
I strongly disagree on this comment just for the record.
Nobody is a fool for praising high quality content from an actual expert in the field.
Calculations are correct and lookback is more than enough. Survivorship bias is an issue, which has been covered here:
th-cam.com/video/OHYExP6RVFU/w-d-xo.html
Do people actually use this? Wow, makes my programs looks somewhat sophisticated lol
I personally don't use it but consider momentum. I am always open for new strategies. Send me yours per mail and I am happy to challenge it 😁
#1 mistake when testing on historical stock data - survivor ship bias, which would drastically affect the performance of a strategy like this; the real-world performance would be around 30% worse than presented in this video
Hi buddy, did you watch the video? I clearly stated that multiple times.
Mate please write a script before you do a video this is so painful to hear
Sorry to hear but I won't write a script. I played around with scripting but people clearly prefer it freestyle as it seems. Besides that I don't have time for setting up a script - it simply isn't worth it.
Thanks for your understanding!
I tried to step by step but I got vaild error.I need some help ! Would you mind solve the problem? What's wrong? Thank you so much.
File ~\anaconda3\lib\site-packages\spyder_kernels\py3compat.py:356 in compat_exec
exec(code, globals, locals)
File c:\users\py
asdaq momentum.py:50
portfolio = mtl.loc['2010-12-31':,get_top][1:2]
File ~\AppData\Roaming\Python\Python39\site-packages\pandas\core\indexing.py:1094 in __getitem__
key = tuple(com.apply_if_callable(x, self.obj) for x in key)
File ~\AppData\Roaming\Python\Python39\site-packages\pandas\core\indexing.py:1094 in
key = tuple(com.apply_if_callable(x, self.obj) for x in key)
File ~\AppData\Roaming\Python\Python39\site-packages\pandas\core\common.py:379 in apply_if_callable
return maybe_callable(obj, **kwargs)
File c:\users\py
asdaq momentum.py:41 in get_top
top_50 = ret_12.loc[date].nlargest(50).index
File ~\AppData\Roaming\Python\Python39\site-packages\pandas\core\indexing.py:1103 in __getitem__
return self._getitem_axis(maybe_callable, axis=axis)
File ~\AppData\Roaming\Python\Python39\site-packages\pandas\core\indexing.py:1330 in _getitem_axis
raise ValueError("Cannot index with multidimensional key")
ValueError: Cannot index with multidimensional key
import yfinance as yf
import pandas as pd
import matplotlib.pyplot as plt
import numpy as np
ticker_df = pd.read_html("en.wikipedia.org/wiki/Nasdaq-100")[4]
tickers = ticker_df.Ticker.to_list()
df = yf.download(tickers,start='2010-01-01')['Adj Close']
print(df)
df = df.dropna(axis=1)
mtl = (df.pct_change() +1)[1:].resample('M').prod()
print(mtl)
def get_rolling_ret(df,n):
return df.rolling(n).apply(np.prod)
ret_12, ret_6, ret_3 = get_rolling_ret(mtl,12),get_rolling_ret(mtl,6) ,get_rolling_ret(mtl,3)
print(ret_12)
print(ret_6)
print(ret_3)
top_50 = ret_12.loc['2010-12-31'].nlargest(50).index
top_30 = ret_6.loc['2010-12-31', top_50].nlargest(30).index
top_10 = ret_3.loc['2010-12-31', top_30].nlargest(10).index
print(top_50)
print(top_30)
print(top_10)
def get_top(date):
top_50 = ret_12.loc[date].nlargest(50).index
top_30 = ret_6.loc[date, top_50].nlargest(30).index
top_10 = ret_3.loc[date, top_30].nlargest(10).index
return top_10
get_top('2010-12-31')
mtl.loc['2010-12-31':][1:2]
portfolio = mtl.loc['2010-12-31':,get_top][1:2]
portfolio.mean(axis=1).values[0]
def pf_performance(date):
portfolio = mtl.loc[date:, get_top(date)][1:2]
return portfolio.mean(axis=1).values[0]
pf_performance('2010-12-31')
mtl.index[:-1]
print(mtl)
returns = []
for date in mtl.index[:-1]:
returns.append(pf_performance(date))
mtl.index[1:]
pd.Series([i-0.01 for i in returns],index=mtl.index[1:]).cumprod().plot()
pd.Series([i-0.01 for i in returns],index=mtl.index[1:]).cumprod()
nas_df = yf.download('IXIC', '2010-01-01')
print(nas_df)
(nas_df['Adj Close'].pct_change() +1).cumprod()
(nas_df['Adj Close'].pct_change() +1).cumprod().plot()
pd.Series([i-0.01 for i in returns],index=mtl.index[1:]).cumprod().plot()
portfolio = mtl.loc['2010-12-31':,get_top][1:2]
portfolio.mean(axis=1).values[0]
def pf_performance(date):
portfolio = mtl.loc[date:, get_top(date)][1:2]
return portfolio.mean(axis=1)
pf_performance('2010-12-31')
mtl.index[:-1]
print(mtl)
returns = []
for date in mtl.index[:-1]:
returns.append(pf_performance(date))
mtl.index[1:]
pd.Series([i-0.01 for i in returns],index=mtl.index[1:]).cumprod().plot()
pd.Series([i-0.01 for i in returns],index=mtl.index[1:]).cumprod()
nas_df = yf.download('IXIC', '2010-01-01')
print(nas_df)
(nas_df['Adj Close'].pct_change() +1).cumprod()
(nas_df['Adj Close'].pct_change() +1).cumprod().plot()
pd.Series([i-0.01 for i in returns],index=mtl.index[1:]).cumprod().plot()