<- Previous Log Select Different Log Next Log ->  
Searching from 2022-11-24 00:00:00 to 2022-11-24 23:59:59.999999.
Query completed in 0.39 seconds
[2022-11-24 06:54:33] --> monrud0lph has joined the channel
[2022-11-24 13:30:25] <Z-Man> If predictions in the wrong direction give increasingly negative accuracy, then that is a sensible value that can be averaged.
[2022-11-24 13:31:31] <Z-Man> There is another caveat: If one week you make a gain of 20%, the next a loss of 20%, those don't cancel out and you end up with a net negative.
[2022-11-24 13:35:31] <Z-Man> The effect is probably not strong enough to matter over all of the other noises you would be getting. If you do want to account for it, the easiest way is to not use the normalization you have, but simply to look at the change of the log of the stock price. Base does not matter.
[2022-11-24 13:36:40] <Z-Man> Say you pick log base 2, then a doubling of the price will be +1 in the log, and halving would be -1. Doubling and halving cancel each other out, as do +1 and -1 if you average.
[2022-11-24 18:49:23] <Lucifer_arma> well, talking about managing gains and losses is an entirely different conversation, but the basic algorithm in my head goes like this:
[2022-11-24 18:50:10] <Lucifer_arma> First, pull all the best predictions out of the database.  Each prediction will get a score (or more than one score) based on overall accuracy of the bot, actual real stock price, and the expected differences between open/close and high/low
[2022-11-24 18:50:56] <Lucifer_arma> This should pull a decent-sized list of stocks that *should* be profitable.  Included in the results here will be some measure of volatility, because that's a way of assessing risk
[2022-11-24 18:52:02] <Lucifer_arma> ok, say this happens as soon as the market opens.  Now the algorithm finds the "best" stocks for the amount of money available and buys them.  It'll determine this by calculating the net gain in actual dollars, because the percentages don't matter that much
[2022-11-24 18:52:45] <Lucifer_arma> then it puts in the sale order, and the price used for the sale order can basically be written as Predicted Price - accuracy of prediction - margin of error
[2022-11-24 18:53:24] <Lucifer_arma> so a stock that starts at $5 and is predicted to rise to $5.50 may be determined to be sold at $0.35.  As long as the price rises to that amount, the sale is guaranteed and the profit is guaranteed
[2022-11-24 18:54:07] <Lucifer_arma> At the same time, a sale order can be placed for, say $4.90, so if the price drops below that, it'll sell.  So maximum loss is $0.10/share, while the gain will be $0.35/share
[2022-11-24 18:54:16] <Lucifer_arma> Multiply that by number of shares and you get total profit
[2022-11-24 18:55:00] <Lucifer_arma> But wait!  There's no "just buy one stock and see what happens" happening.  The whole point of using a computer for all this is because I can put in $100 and have it do this same thing with 10-20 different stocks
[2022-11-24 18:55:49] <Lucifer_arma> Anyway, then we just watch (the computer, that is, does the watching).  If the stock price doesn't get high or low enough to trigger one of the sales, look at tomorrow's prediction and do the contingency plan.
[2022-11-24 18:56:57] <Lucifer_arma> So let's assume that happens and it's tomorrow.  For some reason the stock opens at $5.20.  Go ahead and sell it all and take the smaller gain.  Or, maybe the predicted high will be $6, so go ahead and put in a new sale order (after canceling the other one) for $5.50 and make a bigger gain, but now it took two days.  So average gain is $0.25/day/share
[2022-11-24 18:57:26] <Lucifer_arma> This very basic description of the algorithm will work, provided the predictions themselves are solid to begin with
[2022-11-24 19:00:18] <Lucifer_arma> But the details of the algorithm don't matter if I don't have bots that can make solid predictions.  So, for a bot to make a solid prediction, it must correctly predict the direction the stock price moved, also correctly predict that High > Low.  It also has to come within like 2% accuracy on the difference between open/close prices and High/Low prices.
[2022-11-24 19:00:37] <Lucifer_arma> It's the accuracy of that difference that I want to average together for different stocks.
[2022-11-24 19:03:17] <Lucifer_arma> Meanwhile, as soon as the market closes, download all the new prices and retrain all the bots on each one.  This should be a quick operation, once all the bots have been trained initially on all existing historical data.  The initial training takes about 1-3 minutes per bot, but once it only has to update for new values, that should be extremely fast.
[2022-11-24 19:04:15] <Lucifer_arma> Also, it's worth pointing out that while *one* bot may not get to that 2% accuracy that I want, if, say, four bots averaged together can get it, that's fine, too
[2022-11-24 21:42:47] <-- monrud0lph has quit (Ping timeout: 264 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.

 © NelgTron 2014-2023. Made for . [About this site] [Credits]