Skip to content

Mean-Reversion Trading Strategies in Python Course

November 12, 2020

This post contains affiliate links. An affiliate link means CSSA may receive compensation if you make a purchase through the link, without any extra cost to you. CSSA strives to promote only products and services which provide value to my business and those which I believe could help you, the reader.

In the last post I interviewed Dr. Ernest Chan who is the author of the Mean-Reversion Trading Strategies in Python Course that I will be reviewing in this post. Readers interested in enrolling in the course can follow this link and receive an additional 5% off by using the coupon code: CSSA5

The course is put together by Quantra/QuantInsti which provides algorithmic trading courses in a slick e-learning format for a wide variety of different topics including Momentum Trading Strategies which I covered in a previous post.

Review of Mean-Reversion Trading Strategies in Python

source: Quantra/QuantInsti

A long time ago when I first started trading using quantitative methods, I tried implementing statistical arbitrage strategies with a friend of mine using only an Excel notebook with a data feed. While I had a reasonable understanding of what to do, I didn’t have either the technical knowledge, practical experience, or programming/operational skills to do things properly. While the backtests look great, real-life trading made them look like a mirage in an oasis. Not surprisingly, this “half-assed” operation was a failure. If only I could have gone back in time and taken this course I probably would have had a much better chance of succeeding. The course teaches you both beginner and advanced stat arb techniques and shows you how to work with Python code and connect directly to Interactive Brokers.

The course teacher- Dr. Chan- is not some ivory tower academic or financial economist discussing arbitrage opportunities on paper or the classroom chalkboard, he is an actual hedge fund manager with a successful track record. As a result the course incorporates all the important reality-checks; from the obvious such as transaction costs, to the often overlooked such as considering short-selling availability and borrow costs, to the arcane such as importance of having non-negative weights in index arbitrage to avoid added exposure via stock specific risk. In my opinion this is why quants should strongly consider taking courses on Quantra/QuantInsti if they want to at the very least avoid beginner mistakes that cost a lot more money than the cost of taking a course.

The course starts out defining stationarity and then relates this to a mean-reversion trading strategy. Various statistical tests such as the Augmented-Dickey-Fuller (ADF) test are presented along with both the requisite mathematics as well as code for calculation in Python. Then basic mean-reversion strategies are presented such as those that use Bollinger Bands or Z-scores that can be applied to stationary time series. This is extended to creating an actual portfolio of positions and how to manage this using signals to also calculating P/L. I liked the fact that an explanation was provided for why you would use statistical tests for stationarity prior to testing trading strategies since the majority of traders move immediately to backtesting.

Dr. Chan then discusses the Johansen test as a more versatile test to ADF and shows how to use the eigenvectors calculated for a wide variety of applications. The cool stuff included how to run arbitrage for triplets (3 stocks at a time) to basket arbitrage. Half-life calculations and their applications was also an important part of the content (this is often explained poorly by many sources but not in this case). I liked that the course covered risk management and also how to deal with broken pairs- something that everyone needs to know. Guidance was also provided for the best markets for pair trading. The course concludes with cross-sectional mean-reversion strategies which have been covered a lot on this blog. Python code for everything is covered in this course as well as how to apply turnkey approaches immediately in Interactive Brokers.

Overall if you are looking to get into stat arb whether at the firm level or for yourself this is a great primer. While it doesn’t show you any secret sauce, it gives you all the technical and practical coding knowledge as a foundation to developing your own. So if you are already very experienced this course probably isn’t for you. But if you are experienced but lacking in the technical/practical side you should definitely consider taking the course. There is also a Quantra community to help answer questions and build networks. Hats off to Dr. Chan and the very talented team at Quantra/QuantInsti for putting this course together.

An Interview with Dr. Ernest Chan

November 9, 2020
“A lot of my ideas are just based on logic”

In the last post I reviewed the Momentum Trading Strategies Course by Quantra (a division of QuantInsti) which I reviewed as part of a recent educational journey to improve my quantitative skill set. The next course that I will be reviewing is Mean-Reversion Strategies in Python which is taught by Dr. Ernest Chan. I have personally read Ernie’s book “Machine Trading” which is very well written and full of interesting and practical ideas. I have also been a follower of his very popular blog which was a pioneer in revealing statistical arbitrage strategies such as pairs-trading. Dr. Chan is a thought leader and industry expert and anyone who is in the quantitative field has inevitably come across his work in one form or another. I reached out to interview him to get a few insights into how to think about quantitative models in the modern era. For those that are unfamiliar with Dr. Chan’s work I have provided his very impressive (and extensive) industry and educational credentials below.

Industry Background: Dr. Chan is the Founder of PredictNow.ai, a financial machine learning SaaS, and also the Managing Member of QTS Capital Management, LLC., a commodity pool operator and trading advisor. His primary focus has been on the development of statistical models and advanced computer algorithms to find patterns and trends in large quantities of data. He has applied his expertise in statistical pattern recognition to projects ranging from textual retrieval at IBM Research, mining customer relationship data at Morgan Stanley, and statistical arbitrage trading strategy research at Credit Suisse, Mapleridge Capital Management, and other hedge funds.

Educational Background

Dr. Chan is an industry expert on ‘Algorithmic Options Trading’ and has conducted seminars and lectures on many international forums. Besides being a faculty in QuantInsti, his academic distributions are available on Quantra and on major web portals. Dr. Chan is also an adjunct faculty at Northwestern University’s Master’s in Data Science program. His courses and publications on finance and machine learning can be found at www.epchan.com. Ernie is the author of “Quantitative Trading: How to Build Your Own Algorithmic Trading Business”, “Algorithmic Trading: Winning Strategies and Their Rationale”, and “Machine Trading”, all published by John Wiley & Sons. He maintains a popular blog “Quantitative Trading” at epchan.blogspot.com. Ernie received his PhD. in Physics from Cornell University.

Interview with Dr. Ernest Chan

1) What do you think about traditional factor investing and trading strategies that use technical indicators? Can they be profitable in the modern environment? 

We don’t use machine learning to generate trading signals, but rather to determine the probability of profit of the existing trading signals generated by a basic, traditional quantitative strategy. This strategy can be a factor model or one based on simple technical indicators. This probability of profit can then be used to determine the order size, which can be zero if the probability is too low. 
Factors and technical indicators are still crucial for the basic strategy. I don’t believe that machine learning can replace human intuition and understanding. In fact, it should be used to enhance such understanding and risk management. The input to a machine learning algorithm is nothing but factors and technical indicators.

2) If you could choose between a Momentum and Mean-Reversion approach to trading which would you choose and why?  

I would trade both. Otherwise the portfolio would not be market neutral since momentum strategies are typically short beta while mean reversion strategies are long. Also, momentum strategies are long “gamma” and “vega”, while mean reversion strategies are short. Note that I put quotation marks around such options Greeks because we are not really trading options nor implied volatility. I am using these terms loosely to indicate an increase in tail movements and realized volatility.

3) Why should traders strongly consider using machine learning in their trading versus hand-coding their own quantitative systems or using more simple statistical tools? For traders that aren’t familiar with coding what do you think is the best way to get started?

Traditional quant strategies are too easily replicated by other equally intelligent traders, hence they suffer more rapid alpha decay. ML strategies have so many parameters and nuances that no two traders can possibly have the same strategy. For traders who are not experts in machine learning or programming can start with a no-code machine learning service such as predictnow.ai

4) What do you think is the biggest challenge for newbies trying to design their own machine-learning models?  

Machine learning requires abundant and correctly engineered features as input. I have seen many newbies trying to use 4 or 5 inputs to a ML algorithm. They should instead be using at least 100 inputs.  


5) Do you have a preference in terms of the type of machine learning model you use such as Neural Networks vs KNN or Decision Trees? If so why?

Decision trees, or the more advanced version called random forest, is the preferred ML method for trading. That’s because it doesn’t have as many parameters to fit as a neural network, thus reducing the danger of data snooping bias. Also, the output of a decision tree is a bunch of conditional decision rules, which are much easier to interpret than the nonlinear functions that neural networks use. On the other hand, KNN or logistic regression are too simple – they don’t capture a lot of the nonlinear dependence between different input features and the output return. 

6)  Many traders and market commentators have noticed that markets seem quite a bit different than in the past. The market seems to move much more quickly and reacts to news in ways that are counterintuitive.  Given your vast experience with algorithmic trading what new trends or insights have you gathered in the last few years? Have you made any specific adjustments or recalibrated your models accordingly?

Market patterns often deviate from the “norm” over a short period (e.g. 6 months-1 year), but they often revert to the norm. One needs to diversify so that some strategies are enhanced during such periods, even though others are hurt. Such regime changes can also be detected or predicted to some extent by machine learning.

Thanks Ernie for the interview!

Momentum Trading Strategies Course

October 29, 2020

This post contains affiliate links. An affiliate link means CSSA may receive compensation if you make a purchase through the link, without any extra cost to you. CSSA strives to promote only products and services which provide value to my business and those which I believe could help you, the reader.

One of the biggest barriers to creating a quantitative strategy is knowing how to code. The other barrier is having sufficient theoretical and empirical knowledge. Getting a degree in finance can help with the latter, and a computer science degree can help with the former but if you want to be able to do both you often have to start from scratch which can be very intimidating. I recently took the Momentum Trading Strategies Course by Quantra which is unique because it teaches you both the background theory and empirical research as well as how to code examples in Python– currently the most popular language for algorithmic traders. Given that my coding skills are limited to using Microsoft Excel, this course was especially useful and I even learned a few new things on the research side. Readers interested in enrolling in the course can follow this link and receive an additional 5% off by using the coupon code: CSSA5

quantra.quantinsti.com/course/momentum-trading-strategies?utm_source=david_varadi&utm_medium=affiliate&utm_campaign=mts_course(opens in a new tab)

Note: This course is currently priced at $179 but will return to its normal price of $499 on November 2.

Before getting to my review of the course below it is important for readers to know a little bit more about the service and the players behind the scences:

Quantra is a learning platform for algorithmic trading courses, where through advanced interactive & hands-on learning technology offers content curated by some of the top thought leaders in the domain of algorithmic trading including;

1) Dr Ernest P. Chan

2) Laurent Bernut

3) Dr Terry Benzschawel

4) National Stock Exchange (World’s Biggest Derivatives Exchange)

5) Multi Commodity Exchange (India’s Leading Commodity Exchange)

6) Interactive Brokers

7) Forex Capital Markets (FXCM)

The parent company of Quantra is  QuantInstiwhich was founded by one of India’s biggest HFT firms; iRage, is today one of the world’s most prominent algorithmic & quantitative trading & research institutes with a user base in 180+ countries.

Review On Momentum Trading Strategies Course

First I have to say that this is a really comprehensive course with very slick technology for the e-learning community. The course took me a couple days to complete which was longer than I expected but it also went into far greater depth than I expected as well. To get the most out of the course you should also read the recommended research articles and also work on coding the examples.

It starts off very basic- almost too basic for those familiar with momentum- but gradually builds and gets more advanced with each segment. The topics covered early on include answering what momentum is and why it exists as an anomaly. By the time you get to the fifth section you are being introduced to Python and how to work with commands and loading in data for analysis. You then cover more advanced topics like how to use the Hurst Exponent to cross-sectional arbitrage strategies in futures that exploit roll returns.

In this comprehensive curriculum it seems like every major popular paper on momentum is neatly summarized and the course also covers important topics like Momentum Crashes and risk management. Each segment has examples linked to using Python. There are also multiple choice questions that are there to test your memory and comprehension of the material. As you reach the end of the course you are introduced to even more practical topics like how to automate trading strategies and link to broker APIs.

Overall I was very impressed and I think this is exactly the kind of e-learning alternative that both students and traders/investors need to make their dreams of having their own automated strategy a reality. In subsequent posts I plan to continue to share my learning journey by trying new courses and will provide readers again with a review. Hats off to the team at QuantInsti for being an innovator in this space.

Adaptive VIX Moving Average with Ehlers Alpha Formula

December 4, 2019

In the last post I described a relatively simply method to incorporate the VIX into the well-known AMA or Adaptive Moving Average framework. The alpha formula requires two separate parameters- a short and a long-term constant which requires greater specification by the user. Ideally the fewer parameters you have to specify the better (although it is important to note that logical requirements for maximum lag and minimum responsiveness often tends to dicate the bounds).

Ehlers suggests in his paper on the “Fractal Adaptive Moving Average” using the following formula to compute the alpha in the exponential average:

alpha= exp(- 4.6*(D -1))

where D is analogous to the “VI” or trend factor in the previous post. Note that this formula was adapted in a different way for creating an adaptive volatility measure in this post. I find it to be quite useful and intuitive so it represents a viable alternative to the AMA framework in the last post.

Based on reader feedback I will simplify the percentrank function to be a rolling lookback of 2000 days to make it easier to avoid confusion in replication (the previous used an anchored lookback or “expanding time window”). We can substitute the percentrank of 1/(10-day sma of VIX) for “D” in the formula and calculate the EMA/Exponential moving average of SPY using the alpha output. Here is what the adaptive alpha should look like going back to 1996:

Note that a higher alpha means a faster moving average (less smoothing and lag) while a lower alpha means a slower moving average (more smoothing and lag). We can see in the above chart that the alpha is higher most of the time in bull markets and lower in bear markets (with the 90’s being an exception). No doubt there is a relationship between the VIX and mean-reversion which tended to work well on SPY in periods when the alpha was low. My research in fact shows that profitability was 4x higher buying on down days when this alpha measure was low versus high. What this implies is that you are smoothing more when price is noisy or mean-reverting and smoothing less when price is a decent estimate of tomorrow’s price. Obviously this is exactly what we are looking for when creating an adaptive filter.

Let’s now take a look at how this new transition calculation performs in the adaptive moving average framework. Once again we will compare the strategy of buying when the close of SPY is > 200sma versus the AMA> 200sma.

Consistent with the previous post the AMA strategy is superior to the basic 200sma strategy with fewer trades. The Ehlers alpha method in this case leads to very similar results as using the classic AMA framework for calculating alpha but with even fewer trades. Note that a “D” of 4 vs 4.6 produced a near identical match to the performance and number of trades as the classic AMA framework. In either case I hope this shows the robustness of using the VIX (or you could use historical volatility or GARCH) in an adaptive moving average as a substitute for using the price. In my opinion it is logical to use an adaptive method for smoothing rather than using static smoothing methods or worse yet the actual price in a trend-following strategy.

Adaptive VIX Moving Average

November 26, 2019

One of the challenges with technical or quantitative analysis is to identify strategies that can adapt to different market regimes. The most obvious is a change in the forecast or implied volatility as proxied by the VIX. During more volatile periods we would expect more signal noise and during less volatile periods we would expect less signal noise. But how do we capture this in a strategy? One method is to use the VIX to standardize returns as presented on this blog used “VIX-Adjusted Momentum” in this post. An excellent recent follow-up analysis was done by Justin Czyszczewski which showed that VIX-adjusted trend-following has been recently very successful during these fast moving markets (Tip: using the median or average of O,H,L,C of VIX versus closing data will make the edge of the original strategy more consistent across history). I will show a new variation using this framework very soon in a follow-up post.

Another way to tackle this issue is to vary the lookback length as a function of the VIX. So how would we do that? Enter the basic adaptive moving average framework which seeks to vary the speed or lookback of the moving average as a function of some volatility or trend-strength function.

We can easily substitute the VIX within the “VI” by looking at a standardized measure of how high or low volatility has been relative to past history. If it is higher we want to smooth more or have a longer lookback, and if it is lower we want to have a shorter lookback. This can be accomplished as follows:

Basically we are taking the percentile ranking using all history to date of 1 divided by the current average VIX over the past 10-days. To visualize how this moving average works we can see it applied to the S&P500 (SPY) in the chart below:

Notice that the moving average tracks the price very closely during bull markets and then filters out noise by becoming smoother during corrections. This is exactly what we want to see! Now lets compare a standard 200-day sma strategy that uses the current price versus using the Adaptive VIX Moving Average filter.

The result is a nice boost in gross performance (no transaction costs) that also happens to come with far fewer trades (48 vs 164) which would even further boost net performance. The results are in the table below:

This concept can be extended in several ways including using a simple moving average on the AMA line to create faster crossover strategies that are more responsive to market conditions. Ultimately this is a very simple and intuitive way to adjust standard trend-following as a function of changing volatility regimes.

Which Quant ETFs are Outperforming?

November 18, 2019

One of the challenges of investing in today’s environment is that there are so many investment products that it is hard to keep track. Our recently launched site Investor IQ is designed to help better organize investors and provide them with the analytics required to make better decisions. The biggest risk to chasing performance is if you aren’t on top of what is happening now and instead are chasing the last 3-year or 5-year winning fund (which are likely to mean-revert). Our relative strength rating tracks momentum at shorter intervals in order to better capture future outperformance. Here is a current snapshot of our “Quant ETFs” category under US ETF Digest which makes it easy to see which quant styles are outperforming right now:

As you can see the market is currently favoring stocks that are either undervalued (Value Screen/QVAL and Shareholder Yield/SYD) or exhibit growth at a reasonable price (GARP) which are ranked at the top by relative strength or “RS Score”. This reflects a significant shift away from growth (FFTY) and momentum (QMOM) which are ranked in last place. In fact the signal- which an ensemble of trend and momentum signals- shows that their trends are in a caution or “Hold” position while all other factors remain in a “Buy” signal. The OB/OS oscillator in the far right column is an ensemble of mean-reversion indicators that ranges from 0 to 100 (least to most overbought). Currently all of the ETFs in the list are neutral in terms of timing but nearing overbought territory (>80) where new purchases should either be avoided or existing positions should be trimmed.

The goal of this list was to capture many of the factors that quants use to produce alpha versus the broad index in a long-only format. We deliberately omitted strategies that had the ability to raise cash since I would consider that a separate genre of tactical strategies. If you have suggestions for additional quant ETFs that are not redundant to the members of this current list feel free to leave them in the comments section and I will seek to add them over time pending review.

Investor IQ Website is Live (In Beta)

November 11, 2019

For readers interested in getting signals and analytics on hundreds of ETFs and individual stocks our Investor IQ website is currently live and free during our beta-testing phase. We will be adding new data and analytics gradually over time as well as improving website functionality. The Economic Model is currently hosted on the site and predictions are updated every 2-3 days in real-time. Subscribers will soon receive access to backtests on the economic model (I have received plenty of requests) and other research unique to the website. For the time being we will continue to publish Investor IQ on this blog with limited functionality on a weekly basis for our readers so be sure to sign up to the website as soon as you can!

Current Economic Model Prediction

October 29, 2019

As of the 25th of October the Economic Model changed signal from sideways to bullish. So far the out of sample predictions since its creation have been quite useful for trading during these tricky markets. While Fed day is tomorrow and could change market sentiment, the economic numbers continue to show that we are not nearing recession territory. This prediction is in contrast to many pessimistic news articles that have been published over the last two months. Ultimately news articles focus on very recent data and tend to isolate a few specific indicators that may no longer have much predictive value (or are context dependent). That was the reason for developing an integrated model that uses 50 different indicators along with dozens of derivatives of the same time series within a decision-tree framework. This model will be featured in real-time on our upcoming website.

Mean-Reversion in Trend-Following Performance Using a 120-day Lookback

September 19, 2019

In the last post we showed that trend-following tends to be mean-reverting in the short-term. Data analysis also shows that trend-following has an even stronger mean-reverting effect using a 6-month or 120-day window using the same methodology. Take a look at the chart below using the BarclayHedge SG Trend Index:

In the last post I hypothesized that the mean-reversion effect exists because investors tend to chase recent performance. But there is obviously a lot more going on that drives the performance of trend-followers including lower interest rates, higher correlations, and generally less pronounced trends. Certainly monetary policy worldwide has also played a factor. A very good paper by Spring Valley analyzes some of the factors that have affected trend-following performance. Ultimately the data suggests that you need to incorporate this effect into your strategy or asset allocation methodology in order to be more consistently profitable using a trend-following approach.

Mean-Reversion in Trend-Following Performance

September 18, 2019

In a recent post I showed that the momentum factor has been mean-reverting in the short-term, and that this effect can be used to trade both the factor and momentum strategies effectively. An obvious extension is to see whether trend-following as a factor is also mean-reverting. After all, time-series momentum and momentum have been shown to be related in the research.

To represent the trend-following factor I used the data for the BarclayHedge SG Trend-Following Index which captures the profitability of CTAs that follow a systematic trend-following approach. For consistency, I used the same methodology as the original post: I took the 10-day return and smoothed using a 5-period simple moving average in order to reduce noise. I then took the percentile ranking using all history available at each point-in-time of that smoothed return. Oversold was considered to be when the percentile ranking of the return was below the median (<.5) and overbought was when the percentile ranking was above the median (>.5). Positions are held until the percentile ranking goes back above (below) the median. As a third strategy, I tested avoiding the top quartile of performance (<.75) which is a lower turnover and perhaps a more realistic strategy.

In general we see evidence of mean-reversion in trend-following performance (this is true using a wide variety of parameters) which becomes most pronounced starting in early 2014. The strategy of avoiding the top quartile (long when trend-following has not recently been performing very well) has been effective over longer periods of time.

The lesson highlighted in recent posts has been that to be effective in the modern day environment you have to be a contrarian— buying momentum and trend-following strategies after they have had poor performance. The reason this anomaly likely exists is because people tend to hire and fire managers based on recent performance. As a piece of anectdotal evidence: when I was doing consulting work a long time ago, I used to joke with my colleagues that certain clients were particularly adept at timing when to be in or out of strategies: as soon as they were upset, the strategy probably bottomed, and as soon as they wanted to increase their allocation it had probably topped. After making several poor allocation/timing decisions they usually quit the strategy altogether claiming that it was ineffective.

I made a presentation at a conference in 2010 about mean-reversion in strategy performance using a sample of over a hundred different strategies (including both mean-reversion and trend-following as well as many others). My experience in money management after 2010 has been no different. Unfortunately, all marketing is geared toward recent performance so there is a conflict of interest: if you understand that strategies mean-revert in the short-term then you will be marketing most when clients are least likely to make money. AUM flows tend to reflect that this method works well. It is therefore not surprising that DALBAR studies show that the average investor (most of whom have advisors) significantly underperform their risk benchmarks.