Is Stock Picking Worth It

[ahha]

I would support Yuval Taylor theory if your capital is < $100,000
The most alpha I find in micro-caps but they are extremely risky and require day to day monitoring.
I trade very unliquid stocks (often < 1,000 AD$V). I wait until the order is filled, very often I take profit once there is large, unjustified upspike.

[yuvaltaylor]

I would support Yuval Taylor theory if your capital is < $100,000
The most alpha I find in micro-caps but they are extremely risky and require day to day monitoring.

FWIW, I currently have $1.5 million in microcaps and small caps. 22% in stocks less than $100M market cap; 36% in stocks between $100M and $250M; 28% in stocks between $250M and $500M; 14% in stocks over $500M.

[aebouvin]

I’ve been using the tools here for 6 years and have found a method that consistently outperforms the market. I’ve found that stocks in the $2.5B to $15B market cap range provide plenty of opportunity but don’t have the volatility that the small cap stocks have and they are very easy to trade. I rely mostly on buy rules to identify stocks that have good fundamentals. I haven’t found technical factors to be good predictors of outperformance. The models I create select stocks with ranges for each variable. To do this I first identified cutoff values to create 10 deciles for each variable. The FRANK(variable) will get you most of the way there. You then test each variable to determine how each decile performs both overall and over time. I mostly look at performance relative to the market because my goal is to beat the market. After testing each variable you can then combine the best deciles from each variable and keep testing until you get something that has consistent outperformance over time. It’s a lot of work but since I started using this approach almost 2.5 years ago my annualized return is 36.0% on a 50 stock portfolio. These 50 stocks are rebalanced every three weeks with about 9 stocks being bought and sold so it’s relatively low turnover. I graph my return against the benchmark and calculate the ratio of the two which is the excess return. I create graphs for each decile for each variable too. This excess return has continue to grow even through the Covid crisis. The only time period that it flattened for a few months was last Fall. I’m an actuary and spent my career doing predictive modeling so I have experience with volatile data which the stock market is. The most important thing is to not overfit your models. Combining several factors will get you the outperformance you are looking for. Adding more will just add noise from what I’ve seen. The last piece of advice is to rely more on what works in the last several years. What works in the market changes over time. I expect that someday my models will need to change because the variables are found by others. I’m continually tracking each variable to identify if they are changing. I don’t know of a way of anticipating these changes though.

Thanks all for your thoughts. I was particularly interested in this post by Aebouvin. A 36% CAGR on a 50 midcap stock portfolio over the last 2.5years is nothing short of astounding! A few questions:

1. Is this one 50 stock portfolio or a combination of smaller ones.
2. If the goal is to look for factors that have beaten the market recently (1 year, 2 years, etc.), why bother looking for the factors that contributed to this out performance? Why not just invest in the best returning stocks recently (i.e momentum), as I assume it’s likely that the best performing stocks recently, will have fairly consistent similar factors.
3. Would you care to post a screenshot of your portfolio performance? I am interested in the equity curve shape.

Please only share what you are comfortable with. Thanks all

This is a 50 stock portfolio generated from about 10 separate models. Each of these models generates a list of stocks with an average of between 10 and 30 stocks. There is some overlap with the models so I eliminate duplicates outside of Portfolio123 and then run the portfolio of 50 from an Inlist. The end result are 50 unique stocks. I’ve looked at using momentum and many other technical factors but have not found them to be as predictive as the fundamental factors. There can be periods when momentum can work but through bull and bear markets fundamentals are much better predictors and produce much more consistent outperformance.

Attached is a chart of my live book that’s held 50 stocks for almost 2.5 years.

---

ModelChart.xlsx (110 KB)

[aebouvin]

Erik,

I really like what you are doing. One of the many things I like is that you are effectively testing each factor for statistical significance before you add it to the ranking system.

This should help prevent overfitting.

Combining several factors will get you the outperformance you are looking for. Adding more will just add noise from what I’ve seen.

This seems to be a recognition that the optimization procedure itself can also be a source of overfitting. I believe this is true even if each factor is statistically significant on its own.

As you know (as an actuary) multicollinearity can make the optimization difficult and potentially cause problems with overfitting. Arguably multicollinearity does not reduce the predictive performance of a regression model. But we are not talking about a regression model with P123. We are talking about an optimization method that is prone to overfitting. And also–with the stock market–the correlation of the factors can change. Any statements about regressive predictive models assumes an i.i.d. sample as you know. It also assumes the the sample on which you are making predictions is drawn from the same population.

I assert (without a proof here) that you are exactly right. That one can cause overfitting at P123 when optimizing a bunch of factors—especially if the factors are correlated. I agree that this is true even if each factor is significant by itself.

So here is my question. Have you tried principle component analysis to mitigate the problems of mulicollinearity and overfitting?

The factors in each principle component can be placed into their own node at P123 (with the weights you get from your PCA analysis “factor loadings”). Then the nodes will be orthogonal and can be optimized without the multicollinearity problems.

I am assuming that–with your profession–you have some tools to look into this if it interests you.

In any case, thank you for sharing.

Best,

Jim

I only use buy and sell rules so don’t rely on a ranking system. I have not tried PCA. I have not found a ranking system to be a good method for modeling stocks. Higher or lower for variables isn’t always the best approach. By using the buy and sell rules I’m able to identify stocks with combinations of variables that perform well. This gives me a lot more control than I would have with a ranking system. I think it’s very easy to overfit using a ranking system and therefore to get poor out of sample performance. As I mentioned before my goal is consistent outperformance so it’s important for me to see how the model performs over time with the addition of each variable. Variable selection is very important. There are many variables that don’t work and should not be included.

[Jrinne]

There are many variables that don’t work and should not be included.

Erik,

Thank you. Could not agree more.

Best,

Jim

[charles123]

Thanks Erik for the insight. Very interesting to see the various approaches that people take.