I like it. I would love to play with it and see how I could do with it. can it all run on the same PC with the graphics. or does the program require a lot of horsepower to run. This would be a great way to learn to trade. I could become my own program hahahaha. teaching myself to be the program. lots of patterns. the guy who wrote the program did a great job. I wish I had a job working on this kind of program. I like investigative science. this is really cool. I have a degree in applied math & physics. I love solving these kind of math programming problems. beautiful applications. I never cared as much for the pure theoritical math or physics without real world applications. this is awesome application and awesome advancement in programming and science. I already trade very complex patterns in ES. it would be interesting to see how my patterns match up with your patterns in the sense that as you go long and short what do my patterns tell me. I bet I am already able to do it better than a computer. your patterns are probably more random. my patterns are organized. and not knowing what you have I would have to side with my organization over random patterns. But I could easy be wrong. It would be fun to play with. I really enjoy this kind of scientific research. send me the program and I will split the profits with you. hahaha. I love to trade the ES. awesome market. Is this your program. did you do the work if so you deserve a gold star. great job. I dont know how you picked the patterns or did you just take thousands of random patterns over a long history.
Just patterns I recognize. You have to remember a momentary snapshot has implications or bias. But orderflow counter to that bias can shift the pattern and ultimately another snapshot would indicate the shift. I tried random patterns but couldn’t be trained to hit a high precision or recall. In order to fully apply to random patterns. You would need huge computational capacity taking snapshots every 5 minutes and than using a predictive focal window to tabulate results. The training would have to occur over lengthy timespan. The patterns with low profitability would be discarded. And than you would be left with patterns that make money. But as patterns get arbed out new patterns evolve and emerge. So the program would evolve with time. So essentially trading would only occur with patterns on the timeframes that have the most profitability. So imagine multiple servers running constantly analyzing multiple markets and timeframes. It’s basically acting like AI trader. But if I can think of this and do it. I would imagine it’s already being done 5 years ago. I’m also working on Twitter model to pick direction of SPY. So far the results are very promising. But won’t know for 1-2 years of testing. Thursday and Friday generated sell signals which is pretty rare for it.
The core program has about 70 patterns which I put in myself. Those patterns are just from market experience. I’ve been constantly putting in new patterns looking at results. So the core program is not changed or modified in anyway. So you could say all future images are out of sample .. I tried other subsets of patterns but the profitability dropped substantially. How did I arrive at core 70? I looked at various price structure right before they generated huge price variance. Like double bottom .. double tops .. wedges. So far candlestick patterns have the greatest utility. Since they have more implied information.
Yes this is a great way to go. I certainly wish you all the luck in the world and that you have fun doing it and make a lot of money. which makes it more fun. God Bless you Chewy
What if you wrote an app to take a snapshot of the pattern (whatever it is) that precedes a profitable market move? It would not be hard to quantify the market move in code then you could take snapshots of different time slices that preceded the move. This would decrease the amount of snapshots needed as opposed to constantly taking snapshots via a rolling window.
Yes, you could but what happens is a large variance might be created from random price structure. So the core data set wouldn’t achieve a high predictive potential. You don’t want to pollute the data set, what helped the most was a ‘supervised training’ model. You would need a way to expand the fractal structure to larger and larger timeframes if no meaningful pattern was found in the smaller timeframe. When I filtered the dataset I looked at multiple timeframes. Like a 60 minute double bottom can be seen but nothing on the 1 minute, the focal window for analysis would keep expanding out and couldn’t stay fixed. It implies huge computational power in real time. You could do it on purchased tick data, but I don’t trust the data. I bypassed this by just going through charts in different timeframes and picking the patterns.
It does not have to run in real time. Whatever the case, look forward to seeing how this app works out for your trading. Good Luck!
https://arxiv.org/pdf/1808.00418.pdf The core dataset hit 72% on both precision and recall. https://en.wikipedia.org/wiki/Precision_and_recall
Thank you for sharing your ideas. Some other posters, like Handle123, comgnum among others, at times talked about chart patterns to trade too. It all kind of makes sense to me because I can think of price action is the results of buying and selling pressures, like the forces perturbing a physical system.