IRCLogs for #armagetron in 2022-11-23 • ArmaNelgTron.tk

Searching from 2022-11-23 00:00:00 to 2022-11-23 23:59:59.999999.
Query completed in 0.38 seconds
[2022-11-23 08:37:20] --> monrud0lph has joined the channel
[2022-11-23 11:10:05] <-- monrud0lph has quit (Quit: This place reminds me of Santa's Workshop! Except it smells like mushrooms and everyone looks like they want to hurt me.)
[2022-11-23 11:19:37] --> monrud0lph has joined the channel
[2022-11-23 12:05:17] <Z-Man> Lucifer_arma: Yes, of course you can do that. You just need to be aware of the different ways to average and what they mean. Optimize your average to the outcome you want.
[2022-11-23 12:06:39] <Z-Man> At the very least, you need to normalize the accuracy to the stock value. It's relative errors that count, not absolute ones. It is more useful to correctly predict a raise from 10$ to 15$ than from 100$ to 120$.
[2022-11-23 12:10:17] <Z-Man> I'd simply put into the average what I intend to do with the predictions. Say you plan that you buy 1000$ worth of stock when your system says it is going to rise 10% over the next month, and sell it all after a month. Or when it did rise 10%. Or when it fell 20% instead. Whatever you plan, put it in, simulate what you would do with the AIs predictions, and then your measure is how mch you would make or lose.
[2022-11-23 12:11:26] <Z-Man> (I would not do anything like that at all. It is a good idea, and the problem with that is that someone with bigger pockets, mode data and fatter AI systems probably already tried it)
[2022-11-23 12:12:27] <Z-Man> (If it worked, they are still doing it and you can't compete. If it did not work, your raspii powered AI won't work either.)
[2022-11-23 18:08:01] <Lucifer_arma> Over 80% of transactions on the stock market are already done by AIs by those people with big pockets.  The competition is irrelevant.  :)
[2022-11-23 18:08:34] <Lucifer_arma> I'm still trying to measure how good the bots are at predicting while I'm zeroing in on a more or less final configuration that I'll put into realtime use, so these averages are just a way to measure performance
[2022-11-23 18:09:31] <Lucifer_arma> The basic idea is to predict the lower value stocks that I can afford now over the course of the next 2-5 days, and write a trading algorithm that works on that
[2022-11-23 18:10:15] <Lucifer_arma> so a stock may only increase from $5 to $5.50, but if I trade on that .50/stock/day, I can turn $100 into $120 in a week, and then it snowballs until I *am* that deep pocket person
[2022-11-23 18:11:02] <Lucifer_arma> anyway, yeah, over 80% of stock trades are already automated with AIs, specifically deep learning RNNs, and *that's* why the stock market didn't react to the last couple of presidential elections.  Not because investors liked Trump or Biden.
[2022-11-23 18:12:17] <Lucifer_arma> the specific average I'm looking at, though, is already normalized.  It's a 0-1 average, you know, before you multiply by 100 to turn it into a percentage.  It's the accuracy of the prediction in the spread
[2022-11-23 18:13:14] <Lucifer_arma> so if a stock changed from $5 to $5.50, and my bot predicted it would move only 40 cents, but it predicted the right direction, then the accuracy would be 0.8.  That same bot working on another stock may have the real change of $2.00, but predicted $1, again in the right direction
[2022-11-23 18:13:42] <Lucifer_arma> so that would give 0.5 for accuracy.  So my question is "Can I average the 0.8 and the 0.5 to get an overall measurement of the bot's accuracy?"
[2022-11-23 18:14:23] <Lucifer_arma> I'm only using it to compare different bot configurations (different optimizer, loss, and activation functions, and also different numbers of nodes in the network, and also different decay values for each layer)
[2022-11-23 18:15:14] <Lucifer_arma> in production (and starting in simulation), I'll track accuracy of predictions on a stock by stock basis, but the trading algorithm won't know the accuracy of a current prediction, obviously, because the final price isn't known
[2022-11-23 18:15:33] <Lucifer_arma> but it should be able to look at the trend in prediction accuracy over the last, say, 7 predictions for a given stock
[2022-11-23 21:12:13] <-- monrud0lph has quit (Ping timeout: 252 seconds)
View entire month
DISCLAIMER: These logs of public chat may contain some content which may not be appropriate for all audiences. Use at your own risk.
Logs from 2006-2009 pulled from wrtlprnft
Format changes at: 2015-08-25, 2017-02-20, and 2020-03-23. Times (2015 and later) should be Eastern.
. [About this site] [Credits]