tigerjdengflavortown

I finished up working on the backend hosting for my website, and fixed some weird networking bugs that were causing a failure to fetch market prices for real stocks, and redeployed the website.

Set the whole thing up for shipping by creating a showcase website with mlflow and sqlite on Director4.

I also did some more processing with my 9th training series, and I started work on checkpointing the models so I can test them on a second test range before deploying in real life.

I migrated all the training data from the 9th run series over into the postgresql database on the central workstation I am using. There are some interesting things that I noticed in the results - for example, training with trading slippage does like so bad that it gets worse from the first epoch performance, which is very odd.

I didn’t really get around to logging this devlog for a bit because I was so busy with other work…

In this one I reworked my stock asset list to use lower-absolute-price stocks because I want to deploy it soon, and I can only buy integer shares of stock. I also added a training and testing pipeline to mimic this slippage caused by not being able to get 10-decimal-place exact allocations of stocks.

I ran some training runs using this new data with variations based on whether slippage is used for training.

I completed and transferred over data from my 8th training run where I was trying out different convolutional dilations. The image shows the mlflow graphs of the results and some of my conclusions. I also switched the database backend to use postgresql instead of sqlite (this took a really long time due to like linux directory permissions and stuff) because sqlite was running really slow for all my data.

I analyzed the results from my 6th large training run series and setup the parameters for the 7th.
I also wrote some code to allow for easier sequential training runs of so-called “sweeps” - so I could send one job that ran multiple sweeps for organizational purposes rather than having to manually start them after previous ones completed. At the end, I started the 7th training run series too (image is a gpu status from terminal).

For this devlog, I worked mainly on the system which is used by my school to host sites (Director4), which is where this site is hosted on, so I logged that time under this project. Mainly, I upgraded dependencies that haven’t really been touched in 5-6 years - Python from 3.8 to 3.13, Ubuntu from 20.04 to 24.04, and Django from Django 2 to Django 5.2.

Switched to using mlflow for enhanced logging and visualization. Also started using using torch.compile and tensor-float 32 to make training significantly faster. Sorting out torch compile Dynamo speculation divergence issues took a while.

Finished up implementing a new asset type for Bonds and all the relevant trading, redemption, ui display, etc. mechanics.

Created a specific version of the simulator for public-facing usage.

Was tested and deployed in my school’s math club meeting. Also fixed lots of random bugs which popped up during this attempt.

Refined a lot of things about the website, including user interface, external asset ratio limits, superuser monitoring, etc.

Finished developing trading features such as order executors and flushed everything out.

Deployed to a web hosting service.

Built the whole framework for the website and fixed some random path-based 404 errors.

Finished gigantic training run totaling 256 different models, and compiled/uploaded data from that.
Also fixed by source code to retry after hitting gpu out of memory errors.

Created a pyproject.toml, built application, and added to PyPI and created a GitHub release.

Wrote additional features to the script and pushed to github. Also wrote a readme file.

New version of model with enhanced feature set is being gpu-trained right now.
Will take around 2 days to complete.

I’m working on my first project! This is so exciting. I can’t wait to share more updates as I build.