Month: August 2024

I recently upgraded my OS to Ubuntu 24 (from Ubuntu 22), and was immediately reminded why I always decide afterwards that it was a big mistake. It only happens once every few years, so I forget the pain that happened last time.

Minor issues are fairly easily fixed. Getting 640×480 res on your 4k monitor? No worries, just change the settings back to use the nVidia driver. All your python packages gone AWOL because the system default version of python has been upgraded? Try reinstalling them.

But now there are lots of warnings that there didn’t used to be. ‘This is an externally managed environment, don’t mess with it’. So I could go through the new approach to installing packages system wide, but I’m not sure if I should. Back to using Docker for all my environments. But now jupyter notebooks no longer work. Servers don’t respond. I can’t open a notebook either inside PyCharm or without.

Of course I could go back to using Google Colab for notebooks, or just do everything in .py files. At least they still work. If the packages are installed. Lots of ‘invalid interpreter’ messages on my various PyCharm projects. I have been trying to avoid conflicts by using the system installation wherever possible, but now that that’s been overridden almost nothing works.

When I was doing computer graphics I read somewhere that when a studio starts on a new project they freeze all the software that they’re using, even if the project takes years. One doesn’t want an upgrade on some crucial tool to wreck years of work. I should adopt a policy of getting a new OS only when I get a new computer, which doesn’t happen very often.

Yesterday I had a position open which was down 2%. My intuition was that it would go back up again shortly, but my XGBoost model was giving me a no-go signal. I thought there wasn’t much point in developing the model if I didn’t use it, so I took the loss. For the past day (every six hours) I get another no-go, and indeed the price, while fluctuating a bit, has generally been down. I’m gaining some confidence in the model, especially as so far it gives no-go signals fairly rarely.

I’ve recently been concerned with the problem of understanding the Python language better. I frequently encounter code that I struggle to parse/interpret. Take the following snippet for example:

final_state_locations =
   next_states.flatten(start_dim=1).max(dim=1)[0].eq(0).type(torch.bool)

I’m sure an experienced python programmer won’t see a problem, but I’m basically an amateur. I picked up basic python, then it’s use in a data science context, so mainly pandas, plus a few statistical functions from scikit-learn or similar package. Now I’m learning how to use Python for machine learning and, well, multi-dimensional tensors in PyTorch are a bit harder to work with than a flat dataframe in Pandas.

So, my quest for greater understanding has lead me down an alternative path. Functional Programming. FP is not about understanding code better. It’s about writing code so that it’s more understandable. Less focus on implementation details and more on general intention. There are a few main principles of this programming paradigm, and one of them is immutable state, and functions without side effects. So I’m wondering how this could possibly apply to training an ML model, which is all about changing node weights to minimize some error function.

I guess there is a FP solution to this problem. Instead of changing weighs one creates a new set of weights from the old weights, and then loads the new weights into the model! Replace the old version with a new version instead of changing the old version. I guess that’s how immutability works, which I first came across when learning about the immutability of strings in Java. I’m not really seeing how this is going to make my code easier to understand. I guess I’ve got a way to go in my understanding of functional programming, or perhaps it’s just not a good fit for machine learning. Anyway, it’s interesting I guess.

On reflection, that example I gave at the top of the post is probably an example of functional programming. A new object is created by running some existing data through a string of functions. Not sure where that leaves me.

Perennial question – how much do I need to know to achieve my goals? Especially when I’m not 100% clear on what my goals are.

In all my studies of ML lately I’m frequently confronted by an inadequate understanding of the basic language. I’d like to become much more fluent in Python, but that’s actually a big job. My current reading, on Reinforcement Learning, adopts a ‘create code from scratch’ approach which I appreciate, but that rabbit hole is looking pretty deep.

I’ve reached the part where they are making extensive, and sophisticated, use of iterators and generators to implement certain functions. RL by nature involves very iterative functionality. So, I’m feeling the need to have a better understanding of iterators and generators, and I already have just the course, or courses, on that. Fred Baptiste has some excellent courses on basic Python functionality (basic as in out of the box, not involving third party libraries such as pandas or scikit-learn). However, each of the fours courses is about 40 hours of content. He’s very thorough. But course 2, which includes iterators and generators, definitely builds on material from course 1, especially scope and closures. That’s a lot of material. And all just so that the book I actually want to study is more accessible. And it’s all just a pastime anyway. So what else might I be doing with my life?

I had a strong interest in computer graphics for about 15 years, mostly 3D but also some digital painting, and spent a lot of time doing courses and workshops, and just creating work. Far more than the time I’m now looking at getting to the next level in Python programming. I have another blog documenting that, sadly now rather neglected.

Fortunately I do have a trading system up and running. There’s some satisfaction in that even if it does need a lot more work to improve it. This makes actually putting time into relevant work not so frustrating. Slogging along with no light at the end of the tunnel can be demotivating, but that’s not where I am at the moment.

Whither now? Having got my basic trading system in place I now need to work on the details. I need to actually put together a dataframe (table) with a bunch of returns from several different sources (crypto index, spy, nasdaq, other) all lined up by date. Not difficult, but rather boring. And since my current level of trading is low, and my profits likely to be in the range of a few dollars, motivation is not high. I will get it done but I need to be in the mood for getting boring stuff done, and that doesn’t happen very often. I’m not a details person, which is probably why I suck at languages, art, music, etc.

So I was browsing in my local cheese shop online bookstore and found an interesting looking book on the subject of Reinforcement Learning. I’m trying to stick with PyTorch and am avoiding books that use TensorFlow for code examples, so I went to the GitHub site the authors gave for code access. Lo and behold they eschew all those libraries and basically roll their own. Interesting. One thing I really want to do is improve my mastery of Python as a language, and this seems a good opportunity to do that. It’s great that I can check out what they’re doing in code through GitHub, almost makes up for the inability to actually browse through a shelf of books in a real bookstore like I used to be able to do. Those were the days.

Anyway, Amazon provides a sample to look at before purchase (how I found the GitHub site) and in the introduction they did say one needs a good undergraduate level in probability to follow this book. Probability, not my forte. However, one of the books I have acquired over the years (but not actually read) is Probability for Machine Learning by Jason Brownlee of machinelearningmastery.com. So, as the title says, one thing leads to another.

My main aim is to keep my mind occupied in my doting years. At this stage the trading is just an excuse. I’ve lost too much money in the past through trading that I’ll only risk that again if BTC goes up enough that I have money to burn. At this stage that’s looking pretty unlikely. Even buying a small apartment to mitigate the rental crisis problem here is looking problematic. No matter. I’ll keep those grey cells humming with a good book on probability.

I’ve mentioned meta labelling a couple of times, and I think this actually is an example of it (according to Ernie Chan)

[0.09863082 0.90136915]

These two numbers are the result of running my RL model on the latest price data for the cryptocurrency marketcap (w/o BTC or ETH) index. According to the model, these are the probabilities that not buying is the best action, and that buying is the best action, respectively. This model doesn’t give a definite answer, just the probabilities that each available action is the best one. So this could be used to manage risk. If the probability that buy is the best action is a bit low, say 0.7, one could enter a position with reduced size. Or set a tighter stop loss. Or some other risk management strategy.

Start 11th August 2024. Balance will be updated after each trade is closed. Trades take place at irregular intervals, and for variable lengths of time, so simply numbered here.

UPDATE – STRATEGY CLOSED

I’m moving on to a modified strategy, combining traditional and ML algorithms, one that includes stop losses. Testing shows fewer trades (1 or 2 per month) but fewer, smaller drawdowns. It will be a while before I know how successful that will be.

I don’t have my risk management (trained RL model) in place yet but decided to get started on my ‘trading hobby’ anyway. So, with a prediction from xgboost of an increase of at least 1% in the price of ADAUSDT over the next 5 days, I bought 100USDT worth. I plan to make 100USDT trades, not compounding here. I’ve put a bit more than 100USDT in my Binance spot account to handle any drawdowns. Currently the precision (number of true positives out of all positives) is about 80% so I might lose. I’ll have to consider whether to close when I reach 1% profit, or when 5 days is up, or something else. My strategy is still a bit vague at this point. However this is the first time I’ve traded using a signal generated by an ML model. I’ve seen plenty of warnings not to do that. I guess I’ll find out.

I thought a progress bar might be a nice touch, so I installed the Ultimeter plugin but it doesn’t show for some reason, neither on a page I pasted the code into, nor a post. Maybe my theme doesn’t support it. Too bad. I’ll just report progress from time to time.

UPDATE: Just made my first 1% profit. Was going to hold out for a higher price but my algo predicts the price will not go up by another 1% in next 5 days, so I decided to take profit while it lasted. Be interesting to see what does happen over the next few days.