This discussion is nearly a decade old. But I felt compelled to comment. Statistical arbitrage and market-neutral strategies are common place among retail traders who realized that technical analysis wasn't working for them and decided to do something different.

The number of retail traders who use those strategies is a small minority, because it requires a paradigm shift in thinking and complete abandonment of what they were taught.

Is it a lot of work? Initially, yes, especially if you don't have a background in statistics and coding. But in this era, one can flatten the learning curve tremendously with the advent of Python statistical packages and ChatGPT.

I started my quasi-quant journey back in 2020 when I decided to learn Python. In 2021, I attended a data science boot camp. In 2022, I attended a quantitative/algorithmic trading course. In 2023, I started using ChatGPT the correct way. There have been distractions along the way (e.g., exploring other trading methods). But in 2025, I've returned to StatArb and market-neutral strategies, because I find it as the best fit.

Do you need sizeable amount of capital? Well, it depends. In my opinion, $50,000 to $100,000 in the futures market should suffice for starters, because SPAN margin will allow for margin relief when you take both long and short positions simultaneously. Furthermore, market-neutral strategies are an efficient use of one's capital.

@ZB23 I suspect I'm the only person still around from that thread 10 years ago! The fact that I'm still here, and still doing similar stuff hopefully shows that it does work!

For what it's worth, I would still make both the statements I made back in August 2015. Specifically supporting the points that ...

• The less popular the market, the bigger the chance you'll find an edge.
• Lots of opportunity using multi instrument statistical analysis.

Why do these things work? I think it's two reasons. First what you said "because it requires a paradigm shift in thinking and complete abandonment of what they were taught" and secondly a lot of these edges aren't worth the time and effort of the big boys.

Unfortunately things have got quiet around here and theres a lot less discussion of interesting subjects like this, especially things like this. One thread that has got some replies recently you might be interested in is

Machine Learning Journal Its been a while since I posted anything at NexusFi. Last year, I realized that my rules-based algos will soon be outperformed by smarter ones, probably based on machine learning. So, I am going to teach myself machine learning and convert my algos to …

That ML thread is a interesting dive. I forgot a lot of the ML things that I learned from that data science boot camp four years ago. I briefly enrolled in WQU's free data science lab last year, but had to drop out. I may enroll again in the futures.

Currently, ML is more of a distraction than a help. Good ol' linear regression will suffice for me.

ZB23 

Good ol' linear regression will suffice for me.

Good old OLS. Time tested and reliable (if done correctly).
As they say. KISS, Keep it Simple Stupid!

@SMCJB I'm curious about something. When it comes to hedge ratios, do you use those generated by OLS (for two-legged spreads) and Johansen for (3+ legged spreads), or to you use the conventional 1 lot per leg.

Johansen comes up with some interesting hedge ratios that are more "mean-reverting" than the conventional weights.

(A bit off topic on the topic of this thread, but I wanted to address this topic - that rhymes too!)

SMCJB 

[MENTION=89931] [*]The less popular the market, the bigger the chance you'll find an edge.

This is *very* true.

I started trading my small edge from the ES and YM to Copper (HG).

It works impeccably, no problems, however, *the fills* on larger lot sizes does not work properly because of less volume in HG compared to no problems in fills in the ES and YM due to them being far more liquid.

What's more - I found the same edge works even better on Cotton, Cocoa, Coffee and even on Lumber futures. Why? Because like you said, people don't even trade these instruments or they don't even know that they even exist!

I started to check the available instruments/products list in the prop firms I have accounts with.

Then I started to check my edge in *all* of the instruments listed there and guess what? It's far more easier to trade in markets with less exposure to the general public than to trade the usual suspects, i.e. ES, YM and NQ.

Of course edges come and go and may not work with all, but the instruments which are lesser known, the greater are the chances for edges to work properly.

/Derailing the topic. Thanks.

ZB23 

@SMCJB I'm curious about something. When it comes to hedge ratios, do you use those generated by OLS (for two-legged spreads) and Johansen for (3+ legged spreads), or to you use the conventional 1 lot per leg. Johansen comes up with some interesting hedge ratios that are more "mean-reverting" than the conventional weights.

I use 1 lot per leg, but I would quantify that most of my trading of this type is INTRA-commodity rather than INTER-commodity, so I'm not dealing with the issue of different instruments have different beta's to the underlying market. Specifically I'm looking for over valued and under valued time spreads. The issue here is that volatility, and departure from mean, is a function of distance from spot, or expiration. So the CLZ7-CLZ8 spread will move more than CLZ8-CLZ9 spread. The way I handle this is to have distance from spot, or expiration, as a variable in my analysis[1][2]. I then size positions based upon departure from mean. Then each portfolio is trading the same model across different instruments and I manage the total net delta. So Portfolio 1 may be trading the A-B, A-C, A-D, B-C, B-D & C-D relationships all using the same model, but my tradable output is the total positions across all pairings. One potential problem with this is that A-B and B-C could both be at extremes meaning you would have a bigger position than you might like, because A-C would also be at an extreme.

I've been trading some of these models for over 10 years. Unfortunately I have seen a considerable degradation in performance of the last 2-3 years. Potential explanations of this, are the obvious normal factors. More participants. More data. Less opaqueness/More transparency. But one thing I do wonder about, is whether my activity is distorting the market. I see a lot less major divergences from expectation, especially in less volatile pairings. I wonder if that's because the mere fact that I'm trading these relationships means I'm stopping the divergences from happening before they happen!

[1] It should be noted that if I did use non-1:1 hedge ratio's, the model would give me a distance from spot, or expiration, exposure that it would expect me to hedge! In reality the effect of a 1 day change is so small it is effectively immaterial. In some ways you can view this as a type of carry, and could intentionally construct portfolio's that have a positive carry.
[2] Like most things it's not literally this simple. I'm taking several other variables into consideration, some that would be hedge-able, and some that would not.

I see, in your example, that you have a year of duration between legs. I'm thinking that you embrace the volatility that's involved with that much duration.

ZB23 

I see, in your example, that you have a year of duration between legs. I'm thinking that you embrace the volatility that's involved with that much duration.

I was using the CLZ7-CLZ8 spread as an example. The reality is trading the CLZ7-CLZ8 spread vs the CLZ8-CLZ9 spread (or what we would call the CL Z7-8-9 Butterfly) is a surprisingly volatile relationship. There are a lot more stable relationships available. I have models for the 1 month, 3 month & 6 month spreads. I also have models that compare spreads of different time intervals. While I do trade the 12 month spreads, that tends to be a little less systematic due to the lack of liquidity and higher volatility. You can get in a good position and then it won't trade again for weeks. There are very few markets where you can do something like this. CL has 10 years of listed futures, although the backs trade rarely if ever. The 1, 2 (all) & 3 (HMUZ) month spreads are listed, and quoted out 3 years. The 6 month spreads (M & Z) are listed for the entire 10 years but only the first 4 years/8 spreads are quoted and really only first 3 years/6 spreads trade regularly. The 12 months spreads (Z only) are also listed for the entire 10 years of which the first 7-8 years are often quoted but only the first 5 spreads trade regularly.