I love this analysis, I've seen the current tendency of using the most complex models from the start instead of creating a baseline model to beat. You could name them as "Baseline" or "Naive" approaches, since the word "dumb" may come across as pejorative.
One point to consider is that the S&P500 is not necessarily the correct reference point. Classically, we expect there to be a trade-off between risk and return, so we would expect any model that primarily trades on more risky tickers to outperform the S&P500 relatively frequently. Perhaps (as implied by the Fama-French model) one should compute the average beta, price/book, and market capitalisation of the stocks held by the algorithm on each day during the training period and compare its returns during the test period to a buy and hold strategy on some diverse selection of stocks with similar values.
Thanks for the cool video. One thing I don't quite understand is does this method translate to "sell" as well? And how did you arrive at the Tau value?
I think one of the reasons AI can't trade stock is that the process generating our data isn't a function that follows a stationary distribution, but our model's assume the distribution is stationary.
Maybe i missed it in the video. But how is tau being chosen here? That is a really important part. Also considering all these strategies have different time in the market, one would need to normalize that to be really comparable, no? For example: What if i only use higher decile predictions of the smarter models so far as to match the time in the market of the dumber models? In other words: Let's use the LR Strat but only the highest predictions, so we only have P(Trade)==19%. This way the comparison to D1 (which also has P(Trade)==19%) really makes sense. I would be really interested in the P(Beat SPY) value if you do that:)
I think that if we increase the value of tau for these complex method then we can have P(beat sp) of complex methods similar to that of dumb methods, but at a P(trade) of still around 40-45 percent. Which is a plus for complex methods
Some veteran system traders have found similar results. Really dumb rules like “two days up, one day down” type rules. At least on indexes. It builds from there, but not by much.
Dude... like Your video. What about comparing with SPY for a year return for example? Trading few days in a year You can gain slightly more then SPY doesn't seems increase yearly profit to a lot.. ))) Also a bit question is how much money are losing on a rest days the system is loosing money.. )))
I love this analysis, I've seen the current tendency of using the most complex models from the start instead of creating a baseline model to beat. You could name them as "Baseline" or "Naive" approaches, since the word "dumb" may come across as pejorative.
One point to consider is that the S&P500 is not necessarily the correct reference point. Classically, we expect there to be a trade-off between risk and return, so we would expect any model that primarily trades on more risky tickers to outperform the S&P500 relatively frequently.
Perhaps (as implied by the Fama-French model) one should compute the average beta, price/book, and market capitalisation of the stocks held by the algorithm on each day during the training period and compare its returns during the test period to a buy and hold strategy on some diverse selection of stocks with similar values.
Thanks for the cool video. One thing I don't quite understand is does this method translate to "sell" as well? And how did you arrive at the Tau value?
I think one of the reasons AI can't trade stock is that the process generating our data isn't a function that follows a stationary distribution, but our model's assume the distribution is stationary.
Maybe i missed it in the video. But how is tau being chosen here? That is a really important part.
Also considering all these strategies have different time in the market, one would need to normalize that to be really comparable, no?
For example:
What if i only use higher decile predictions of the smarter models so far as to match the time in the market of the dumber models?
In other words: Let's use the LR Strat but only the highest predictions, so we only have P(Trade)==19%. This way the comparison to D1 (which also has P(Trade)==19%) really makes sense.
I would be really interested in the P(Beat SPY) value if you do that:)
Pretty valid points, I was going to ask the same questions.
I think he explained the idea in the ROC curve video.
Good point
How do u choose an appropriate tau?
I think that if we increase the value of tau for these complex method then we can have P(beat sp) of complex methods similar to that of dumb methods, but at a P(trade) of still around 40-45 percent. Which is a plus for complex methods
Don't forget the control "dumb" technique: buy and hold.
Some veteran system traders have found similar results. Really dumb rules like “two days up, one day down” type rules. At least on indexes. It builds from there, but not by much.
Can you do overview on the more crazy models please
How would you reconcile this with the common wisdom of buying and holding? Can that be added as another dumb method?
What does Tau mean in this context?
Hey i am preparing for time series, is your videos will be enough to learn time series.
Dude... like Your video. What about comparing with SPY for a year return for example? Trading few days in a year You can gain slightly more then SPY doesn't seems increase yearly profit to a lot.. ))) Also a bit question is how much money are losing on a rest days the system is loosing money.. )))