Post New

https://robotwealth.com/brave-new-backtest/

I was going to quote the article but it starts to exceed "fair use", so just read it.

The article assumes on is using AI in the form of LLM’s but this isn’t at all indicative of using a mixture of fundamental and momentum based machine learning to construct predictive models.

My “first” somewhat primitive ML based system trained on data for 9/14/2015 till 9/14/21 and has been running out of sample data from 1/19/22 till the present holding 25 stocks with no retraining.
The SP500 is up 45% in that period the simulation is up 116%. This is with a 163% annual turnover.
This only used 46 mostly fundamental features but used an ensemble of different machine learning models including both tree and linear based regression and different buy and sell criteria based on the future predictions.

I have better current models but have not yet invested significant sums in the ML models yet as I’m one of those older retirees who logically should minimize the risk factors. But one has to have some challenges in life. But I am absolutely convinced the tools available are a very big improvement for analyzing factors for their performance and how they interact.

Yes, I'm using AI but to do calculations. The "top 1, 5, 10 1-year momentum stocks in the Nasdaq100 traded monthly, with a .1% spread" that I analyzed using Robbie's tester I am testing now. I'll have an extra year and more data later. Initially:
Average results across all 21 start dates to 1994; "timing" means out of the market when SPY is below SMA325:

Strategy	CAGR	Beta 	Sharpe	Sortino	Max DD	Max DD recovery	std dev	turnover Treynor
Untimed Top 1	32.7%	1.40	0.79	1.13	-80.1%	529 days	55.3%	9.26x	0.312
Untimed Top 5	22.0%	1.26	0.74	1.03	-77.1%	3826 days	35.1%	7.78x	0.208
Untimed Top 10	22.0%	1.21	0.80	1.10	-62.3%	1192 days	31.0%	6.66x	0.204
Timed Top 1	30.4%	0.91	0.79	1.02	-71.5%	385 days	48.3%	7.54x	0.418
Timed Top 5	22.4%	0.85	0.81	1.00	-53.2%	1868 days	30.9%	6.61x	0.295
Timed Top 10	21.3%	0.81	0.85	1.04	-42.5%	1867 days	27.1%	5.87x	0.285 

Nasdaq100 had a 14 year period to reach a new high so timed strategy was 1/2 that with better MDD.

More later.

So yes I'm not saying AI is a waste of time, just you need to use it wisely.

Strategy            CAGR      Beta      Sharpe    Sortino   Max DD    Max DD recovery     std dev   turnover Treynor
Untimed Top 10      22.0%     1.21      0.80      1.10      -62.3%    1192 days            31.0%     6.66x     0.204
...
Timed Top 10        21.3%     0.81      0.85      1.04      -42.5%    1867 days            27.1%     5.87x     0.285 

Interesting that timed had lower Sortino and longer recovery. That seems odd.

The 1 and 5 stock versions had quicker recoveries. I haven't looked at detail yet so maybe there were some whopper losers in positions 6-10 requiring longer recovery. Also the stats I posted were times to reach new highs; I didn't look to see if there were many that came close so it is possible some of the strategies were mostly recovered and then flat without hitting a new high quite.

Great post.

Whenever I get anything from AI, I always get this spidey sense that all may not be as tidily wrapped-up as my AI printout implies. Longmore’s post sums it up nicely.

Thanks.

musselmant, thanks for your work on the Nasdaq100 stocks. If its not much trouble, could you show how your results would compare to QLD and TQQQ? With timing and without?

Post New

Mechanical Investing FAQ

---

Contact Shrewd'm
Contact the developer of these message boards.