Offshore

Moon Dev
Renting Wall Street Power For $1/Hour: Use Cloud GPUs And Python To Automate Your Trading Edge

no more paid courses needed because the actual secrets to building an edge are hidden in sight within the code you write yourself instead of some guru’s outdated pdf. most people think using ai in trading is just asking a chatbot where bitcoin is going next but they are looking at the wrong data entirely.

my name is moon dev and i believe that code is the great equalizer because through losing money with liquidations and over trading i knew i had to automate my trading. in the past i spent hundreds of thousands on devs for apps thinking i would not be able to code myself but with bots you must iterate to success.

i decided to learn live on youtube and now we are here with fully automated systems trading for me instead of getting liquidated. if you have ever felt that punch in the gut when a position gets wiped out you know why i am obsessed with removing the human heart from the execution.

to actually win in this game you have to follow a process called the rbi system which stands for research backtest and implement. most beginners fail because they try to build a bot immediately without knowing if their strategy even has a statistical edge in the historical data.

the r is for research where you look at academic papers on google scholar or listen to veteran quants on podcasts to find ideas that have worked for decades. once you have a basket of ideas you move to the b which is backtesting to see how that idea would have performed over the last few years.

if the backtest shows you a profit only then do you move to the i for implementation with tiny size to see if it works in the live market. you are probably wondering why everyone else is looking at the same open high low close volume data while getting chopped up every single day.

the real alpha is hidden in unique data specifically liquidation data because when traders get forced out of their positions it creates a massive vacuum that price often fills. everyone else is looking at what already happened while we are looking at the forced selling and buying that is about to drive the next move.

this is where the machine learning comes in because we use something called an lstm or long short term memory neural network to find patterns in these sequences. lstms are specialized brains for numerical time series data which is why they are much better at predicting price moves than something like chatgpt.

chatbots are great for writing code or chatting but they suck at predicting numbers because they were not built for time series data. a transformer model like the one inside cursor or claude looks at text context while our lstm looks at the relationship between liquidation spikes and five minute price targets.

you might be tempted to run these models on your local computer but you will quickly realize that training on 17 million records can take 25 hours or more. i realized that as a trader i am making a trade of either dollars or time when i decide where to train my models.

you can actually rent high end gpus in the cloud like an rtx 4090 or an a100 for less than a dollar an hour and get 20 hours of your life back. training a model 10 times faster than your home computer means you can iterate and find winning strategies while the rest of the world is still waiting for their progress bar.

one of the biggest roadblocks you will hit is when your data has infinite values because some altcoin prices are so close to zero they break the math. we found nearly 8000 infinite targets in our data set which would have caused an infinite loss during training and crashed the entire model.

instead of just deleting that data we implemented a capping strategy to keep the math sane while preserving the raw alpha of those small cap liquidations. we do not want to discriminate against small coins because a 100k liquidation on a tiny asset can still be a massive signal for the broader market.
[...]

1 view12:10