Weekly Algorithm Review: 10/21/2023 to 10/27/2023
Performance Rankings
Sector Neutral: +0.62%
Market Neutral: +0.49%
Variable Sector Neutral: -0.54%
Variable Market Neutral: -0.85%
Long Term Portfolio: -2.04%
Overall Market: -2.36%
Base Algorithm: -2.59%
Another fairly solid week for the portfolio. It beat the market fairly handily. Most of that is from its emphasis on small market cap stocks (it essentially tied with the equally weighted market etf), but a win is a win. Furthermore, our hedging algorithms all saw substantial out-performance this week. We’re fairy bearish on the market these days, so it’s important to us that we see strong performance there.
The Base Algorithm however, has given us an under-performance this week.
The distribution of returns by ticker is similar to what we saw last week. The algorithm dodged the biggest losers, but simultaneously, it missed out on the biggest winners.
Sector distributions are also similar to what we saw last week. Again, the algorithm leaned bullish relative to the portfolio, and again, it got punished for it. Under different circumstances, I might want to keep watching this iteration of the algorithm. An algorithm leaning bullish isn’t the worst thing right now. We’re fairly bearish on the market these days, so were we actively investing in this system, we would probably be using one of the hedgers. And over the 3 weeks of the beta, the market has gone down as we expected, and consequently all of the hedgers have out-performed by large margins. That said, a strong Base Algorithm is our primarily goal, even if we wouldn’t necessarily be using it right now.
At this point, we’re 3 weeks into the beta release of this algorithm. We’ve had 2 weeks of poor performance, and 1 of okay performance. This is a shorter amount of time than what I’d like to do for a paper trade, but at this point we are making the call to end this beta release. This algorithm will not be seeing a full release. As such, I will be cutting its analysis a bit short tonight. It won’t do us much good to deep dive a system that, in a few hours time, we will completely retire.
Instead, I’ve gone through some of the more promising algorithms from our recent round of development, and put them through some additional testing. During our last round of development, we tested algorithms using a portfolio comprised of the 200 highest average dollar volume stocks from the S&P 500 (as of January 1st 2023), with data ranging from January 1st to September 15th of this year - with a focus on everything after July 1st, when our problems with the previous algorithm began.
With some time passed from the beginning of that round, we now have extra data we can test these same systems on. I’ve taken our most promising systems (based on the initial dataset) and given them an additional round of testing: on data ranging from September 16th to October 24th.This was data not considered in our initial round of development, and can therefore work as a testing data set (albeit a small one). Our findings are as follows:
The reworked alpha signals have worked against us. Algorithms that utilize them, and appeared promising in our initial data set have all performed poorly in this testing data set. This is strong evidence that the signals chosen were the result of over-fitting to our training data. As such, we will be reverting to the original alpha signals.
Giving more weight to older days in the algorithm’s training data has improved its performance. I suspect that, some time in the future, this parameter will be changed again. The market has phases on this. At times, things are changing quickly, and it’s important to focus on recent data. At other times, it’s worth using older data if it means more data train on. Earlier this year, our systems did fairly well looking only at recent data. But in more recent training data, utilizing older information has paid off well. This is a change we’ll be keeping.
Anti-trend-following features helped us in backtests on the S&P sample, but lost us money on backtests with our long term portfolio. This was the result I found most surprising. We’ve tried to resist trend-chasing before, unsuccessfully. Even though our current solution to this problem is more involved than our previous attempts. I’m not surprised that it didn’t work. Rather, I’m surprised that it worked on the S&P sample, but hurt us on our own portfolio. It’s hard to draw large scale conclusions from such a short sample period, but tentatively, I think the fundamental-based analysis that goes into our portfolio is the reason for this. Our goal in creating our portfolios is slow, stable growth of every stock in it. Even if we don’t always hit that target, we hit often enough to shift the odds here. For now, this is being reverted - but it will definitely be looked into again in the future.
That leaves us with our newest system. It’s fairly similar to our old one, but it gives significantly more weight to older days when training itself. We’ll be starting a second beta this week, with this system instead.
Before said week begins, I want to clarify: backtests are not paper trades. I have more confidence in this system than I did when we released our previous system, but it’s more important to us that we do things right than that we things fast. This system will be going through a beta release just like our last one did. As always, it’s possible that our testing data isn’t representative of the market conditions coming in the future. We have high standards for our releases, and in order for us to have the confidence to move this system to a full release, we’re going to have to see successful paper trading from it. As always, I will give updates on our sentiments and decision making process as the beta proceeds.
Misc. Data For The Week
