AI musings

When I hedged the dollar I did this by buying eur futures in my manual/special/junk account. Of course that account did very well since then, but I’m happy to say that the other accounts have recovered as well, i.e., in euros they are above the February levels. Lately the asset allocation portfolios (most of my money lives there) have joined the party, while trend following and the day trading strategies have been doing well for longer. The asset allocation portfolios get hurt when there is elevated, but not extreme, volatility. The dollar used to provide a cushion in these times but in the last round the dollar was the cause of the problems. I’m not sure what to do with it. For now I’ll keep the hedge because the dollar seems to be deliberately weakened. Still, in the next market correction everybody might be screaming for dollars again. You can’t ignore it is the currency of the largest economy. My trading might be systematic, but these are still choices I have to make and they keep me up at night.

My intraday program can now mix mini and micro futures. For instance if I need to buy 2.3 NQ futures, that would be 2 NQ’s and 3 MNQ’s. On top of that multiple strategies can trade the same contracts and I have a nice visual way to represent the positions. It was harder than I initially thought and at one point I threw away everything because it became way too messy. Taking a step back gave me an idea for a clean solution that will benefit me in other ways as well: I’m also working on strategies based on cointegration. To facilitate this I wanted to develop a specialized backtesting engine. Yak is used for asset allocation backtests, Bison for intraday technical analysis based strategies. It seemed fitting to add another bovine to the zoo: the Buffalo.

At about the same time as I got this idea, I upgraded my AI assistent to an AI agent and was very impressed with what it could do. It made me wonder if it was possible to develop an entire program with the help of the AI agent. Long story short: not yet. Initially I liked the idea of being promoted to project manager instead of a humble programmer. It was like having someone code for you, but instead of sitting down in the morning, explaining what should be done and then waiting for a day to check how far they got, the AI would be finished in 5 minutes. Initially things were progressing very well, and quickly glancing over the resulting code it looked like a job well done. But when the program was in such a state that the first tests could be done it was clear there were all sorts of bugs. Despite being very optimistic (the AI’s seem borderline ADHD’ers) it could not pinpoint the cause of the bugs. Let me give an example: one task would be to load data of NQ, ES, YM and RTY futures, starting in 2025. For some reason the NQ data did not start in January but later, in March. As a human, it quickly became apparent to me that the datafiles weren’t loaded in order of ascending dates. The AI made the strangest additions to the code to either understand the problem or make it go away, but to no avail. A couple of lessons here: if your AI can’t do what you want it to do: start over or do it yourself. Also, don’t put it under pressure. It seems to make less mistakes when you are gentle with it.

The AI’s may not be ready yet, the progress is remarkable. Right now, we are still needed for the ideas and the analysis but this is quickly changing. Then again, in 2015 I remember saying that in 2020 all new cars would be fully autonomous. It’s now 2025 and I’ve just encouraged my youngest son to get a drivers licence. Oh, and I was the one that said that google was overpriced at its IPO. I’m terrible at predicting the future.

A new addition to my zoo

For several years already I have made weekly posts on a forum on things that keep me busy related to trading. It might be a good idea to do those posts here from now on. Anyway, for people on that forum the title of this post will be less cryptic than for others that don’t know me. Most of my programs are named after animals. I’ve gathered quite a collection already:

  • Dolphin, my real-time risk management system
  • Octopus, a trade logger
  • Woodpecker, a real-time trading engine
  • Yak, a simulator / backtester for asset allocation strategies
  • Bison, a backtester for intraday strategies
  • Swordfish, a dartboard for manual trading
  • Aardvark, a trading strategy
  • Kangaroo, a trading strategy
  • Elephant, a data gathering tool
  • Saint Bernard, a helpful dog that keeps track of running programs

There are others, like Donkey, Axolotl, Tarantula but those are largely defunct. Currently I’m working on Buffalo, which I thought was a fitting name for a new backtester/explorer for strategies based on cointegration. Most of my programs have the same look and feel (if they are gui programs) and that’s because I usually start developing a new project by issuing a single command:

./createProject.sh <project name>

This sets up a complete directory structure for the new project and fills it with relevant class files (I’m using java), and puts them under version control. Automatically I have a program that has the basics of a gui (that remembers its location and size) or headless application, command line handling, license processing, (remote) logging, configuration settings, splash screens, and help/about forms.

It’s pretty cool, because this is always tedious work which is now done with just this single line. The only thing I have to do afterwards is add an image for the splash screen, an icon and a function that returns a description of the projects purpose.

Normally after this my work really starts, but this time I’m working with an AI agent. So far the results of this collaboration have been mixed: the AI is often surprisingly good at making code, and sometimes it doesn’t seem to understand some basic programming structures. What I’ve learned so far is that if the AI doesn’t get it right in three or four times, don’t try to push it further. It will only add more bloat. Better to debug and fix the program yourself.

I like programming, it can be challenging but also very rewarding. In a sense it’s a pity the AI’s are on the rise because they take over so much of the work that I used to enjoy. So why use them? Because they also take over much more work that is boring. There seems to be no middle ground here: either you use the AI or you don’t. For this project I’ve chosen to use one. It feels like I’ve been promoted from programmer to project manager, because that’s what the work has been transformed into. I’ll let you know when or if I have demoted myself.

Protected Scalping

Scalping has always intrigued me. If you consider that the market can go up, down or sideways then by buying a stock you really bet on only one of those outcomes: it should go up. By continuously buying and selling for a few ticks higher you bet on the market to go sideways, but if it goes up or down by a relatively small amount you still make money. The real problem with this approach is that your losses mount exponentially, while your profits grow only linearly: if you start buying 100 lots at every whole dollar, start at 100 and the stock goes to 90 by the end of the day without going up enough to trigger any sell, you will buy at 100, 99, 98, etc. By the time the price hits 90 you will have bought 1000 lots at an average price of 95. That’s a loss of $5000, while you will only make $100 per buy-sell round. That’s a very bad risk-reward situation.

So what if we hedge any positions of significance with puts and calls? According to options theory that should not bring us any money because of the no-arbitrage argument. Still, I’m curious and that’s what I set out to do.

Here’s the basic idea: if the current price is 100, then I want to enter buy orders at 99, 98 and 97, and sell orders at 101, 102, and 103. The space between those orders (1.00 in this case) can be specified and will be called tickSpread. The number of buy and sell orders is N. The order size is aptly named orderSize. At all times I want my position in the stock plus N*orderSize to be smaller than 100 times the number of puts I’ve got. If that’s not the case I first need to buy enough puts. Similarly, ,minus my position in the stock plus N*orderSize must be smaller than 100 times the number of calls I’ve got.

If those two conditions are satisfied, we actually place the buy and sell orders in the stock. If an order is executed, we place a new order of the opposite side one tickSpread below the order price. For example, when a buy at 99 is executed we add a new sell order at 100. We keep playing this game, all the while ensuring we have enough puts and calls.

Is this scheme going to make any money? Well, the good part is that the price runs away from our starting point we are protected by the options. Roughly speaking, the loss now grows linearly with how far the price moves away. The bad part is twofold: the position in the stock can still become very big. In Europe we can’t do portfolio margin, meaning the broker/clearer doesn’t net the risk from the options plus the stocks but considers them individually. In other words: you will need a lot of capital.

The other bad part is the decay in option value. Options can be tricky, and as I’ve stated earlier, from the theory you should not expect there is any money to be made here. The options are purchased at the worst possible time: if the market is going down you will need puts, and those will be expensive then. If the market is going up right after that, you will profit by selling portions of stock but alas, the puts will lose value. But there is hope: if the stock then goes down again to the same level, you won’t have to buy new puts because you already have enough of them. But you will make money on scalping the stock!

I’m a bit skeptical this will turn into anything useful, but at the same time it seems like a fun experiment. Normally I’m not a fan of paper trading, but in this case I have absolutely no idea how much of a disaster this can turn into. In the coming days I’ll post some statistics on trading NVidia using this method. Here’s the first one. It present the numbers at 20:22 CET (1.5 hours before the close). TickSpread is 0.20, orderSize = 50. This means a round trip scalp will make $10, minus 2 times $1 transaction costs = $8.

2025-01-10

Start position:NVDA: 0
Start price:135.60
PnL
End position:NVDA: -350183
Feb21’25 135 call: 3 @8.76124
Feb21’25 136 call: 3 @8.528-68
Feb21’25 137 call: 3 @8.611-240
Feb21’25 134 put: 3 @7.561-303
Todays protection costs:10038.30
Trading range:134.20137.20
Total:-404
End price:136.30
#trades:2400 buys @ 135.9172750 sells @ 136.065

That doesn’t look to fancy. The cost of protection is now 10038 and we need to make that back by scalping $8 per buy/sell pair. We do have tremendous turnover though: we bought 2400 lots, which translates to 2400/50 = 48 trades. If the market would magically return to the starting price we would have made 48*$8 = $384. If that’s an indication of what we make per day we would need 26 days to pay for the insurance. Hmm. 26 days from now is somewhere around expiration day. That would mean the options are priced amazingly accurate. It is very likely NVidia will move beyond 134 – 137 and that would mean more option buying. Stay tuned!

The permanent portfolio loves volatility

The drop in equity markets in Q4 of 2018 has hurt a lot of peoples portfolios. Strangely enough, the permanent portfolio loves these periods of high volatility. In itself it will not decrease in value dramatically because that’s what it is designed to do: when markets are nervous, money flows from equity to gold and/or bonds. When things return to normal money flows back but for some reason you end up with a higher value for your permanent portfolio. Even without rebalancing. I have seen this phenomena many times and my best explanation is that in the first round gold and bonds are bid up by the so called weak hands fleeing out of stocks, and in the second round additional money flows from smart money, that was parked on the sidelines, into stocks again.

Whatever the cause, I’ve now seen this happening so many times that I rejoice when markets drop. It doesn’t make me a popular guy at parties but I’ll accept that!

Hello world!

So here it is, my first post on this blog. First of all, thank you for visiting. If you didn’t get here by chance we are probably connected, sharing some similar interests or views. That’s also the reason for putting up this blog: I would love to get in touch with like-minded spirits.

Right now the content is very limited, but I have plans! I’ll discuss my own flavors of great portfolios, will have a couple of things to say about MIFID and PRIIPS, or regulation in general, add a couple of games from behavioral science and will write about my projects du jour.

Feel free to join the discussion in the comments. If you want to reach out to me personally, email smeets <at> marksmeets <dot> nl.