Algorithmic Week In Review: 08/14/2022
Current Algorithms - Recent Performance
The last week has been similar to our usual expectations. I really want to shine a light on our current experiment: FeelGoodInc, because its debut week has gone quite well, and the reason will help shine a light on why some of our systems have underperformed.
The biggest anomaly we’ve seen is that, more often than expected, the base algorithm has outperformed both of our hedging systems on days the market was down. This might seem counter-intuitive; that’s because it is. When the market is bearish, we expect a high-beta system to underperform a near-zero-beta system almost every time, but that isn’t the case.
The more I look into it, the more my initial hypothesis seems correct: the base algorithm is just better at modeling trading signals. The base algorithm is trading raw price (it doesn’t hedge at all), so all of its trading signals are built accordingly, using raw price information.
Take Kokomo as a counter-example. It hedges its long positions by trading the market as a whole. To account for this in the modeling stage, it calculates its trading signals by taking raw price information, and isolating it from outside market influences. In theory, this should improve its modeling, since it’s creating trading signals based on the value of what it will be holding, but in practice this hasn’t been the case. People don’t trade based on RSI-minus-market-influence nearly as often as they trade on RSI. For this reason, even if they’re looking at more isolated info, a buy signal from the hedgers has been less consistent than we expect to see.
FeelGoodInc is our solution. It combines sector-based hedging with trading signals calculated on raw price information. This has improved its backtest results substantially, and it’s tentatively holding up in live testing.
Over the first week we’ve watched it, it’s only really had one bad day, that being Friday. On that day, it had a negative return, whereas both the market and GoodVibrations (our other sector-hedger) did positive. That aside, it consistently either beats the market, or is positive. In a system optimized for Sharpe ratio, that’s a really strong sign.
I’m interested in advancing FeelGoodInc to the next stage of development: a beta release. I’ll be talking to Allen about this later, but if opinions are positive, our platinum users can expect to see publications in the experimental channel.
New Developments
As some of you may be aware, our trading bot stopped posting notifications at about 11 AM CT on Friday. This is because we’ve been running it on my personal laptop, which crashed around then.
We’re working on moving it to an Ubuntu server, which should be much more reliable. This, however, has required me to become acquainted with Linux. I ask that you all be patient with the bot if it has some bugs in the next few days. In the long run, this should drastically improve reliability.
We are going to be dropping support for the neural net system. I did a comprehensive backtest over the weekend, and its performance isn’t up to par with the Base Algo.
Note that this is NOT tomorrow’s exposures; this is their exposures on Friday. The neural net has consistently had a higher or similar beta to all sectors, including the market as a whole. It’s also a very blackbox system. If the Base Algo recommends a questionable move, we can figure out exactly why. With the neural net, we have a harder time giving that level of explanation. Combine this with its lower historical Sharpe than the Base, and we get a system that takes more of the same risks, but makes less money. I’m content to say this experiment has failed, and we will no longer be publishing NN reports anywhere.
That’s all for tonight. Thanks for reading, and good luck this week. Let’s make some money. Here are our tentative exposures for tomorrow:
-Asher
