11–20 of 178 posts
Pretty underwhelming. From the very limited experiment demonstrated, the LSTM seems to be playing catch up rather than actually predicting trends.
You say that because the prediction line lags the price line? That’s not playing catch up, if the prediction was made at time 0.
I wonder if someone could comment on how AI is used at big investment banks? I assume its much more than technical analysis (like this model). I imagine that NLP might be helpful for quickly ingesting a news feed and then making buy/sell decisions based on the news in less time than humans can read and react?
Other examples include analyzing the words/sentiments in earning calls, to see if certain words indicate a bearish or bullish signal. For example if a CEO mentions the word "headwinds", does that historically lead to a much better quarter next time?
Some other companies do analysis with public/private data. For example, you can use a weather dataset to calculate the average temperature for all cities where there are restaurants owned by Cheesecake Factory in a quarter, and see if it will negatively or positively impact sales.
Working out a function that passes a split test is inevitable and easy tbh. That function then making money is highly unlikely. This is especially true if your data sources are public.
The problem with these types of systems are it's difficult to do backtesting on a single continuous data stream... I write about that here: https://blog.projectpiglet.com/2018/01/perils-of-backtesting... I recommend doing something similar to the original post. Neural networks tend to produce valid looking output for a stock price easily, a random walk does too. You have to find correlations, causations, and then rev…
I do not see enough data to backup the claims of how well it works. And if you can't provide that - what good is it as a tool?
Imagine having the whole Twitter stream, maybe categorized by some network in an automatic way.
Also prequential evaluation.
I preliminarily did the same thing as in this article a while back (with crypto data, since it's easily accessible), just as a way to learn keras. Somehow my network always learned to output a delay of the input no matter how hard I tried to shape it. I've searched through literature briefly and some examples. Some blog posts even claimed they had good time series prediction when I clearly saw they were having the sa…
Pretty underwhelming. From the very limited experiment demonstrated, the LSTM seems to be playing catch up rather than actually predicting trends.
Many data science thought pieces don't quantify the results, and it's suspicious that they're not included in detail (typically, the R^2 of a simple stock market prediction model is super low, making it impractical putting actual money on the line as it's barely better than guessing randomly: https://twitter.com/minimaxir/status/1021885939361042432 )