Comment by pyrrhotech | Hacker News Reader

Comment by pyrrhotech | Hacker News Reader

Parent Full thread

This is different, no martingale here and no leverage used at all in obtaining the returns, the models are always either hedged (effectively in cash or 0% long) or unhedged (100% long). This is also not a high frequency trading system. Most models have an average holding period of about 2-4 weeks, though with high variation depending on market volatility levels (has been much shorter than that this year). Each detail page gives a brief, high-level overview of some of the indicators the model uses to detect trend strength, trend reversals and mean reversion opportunities that ultimately produce a signal change.

People often think beating the market is impossible, and for an unaided human, it may be. But just like Magnus Carlsen cannot come close to defeating Stockfish at chess, sophisticated algorithms that analyze far more data than possible for a human are on an entirely different playing field. The ironic thing is the growth in the belief in the EMH and passive investing has actually made beating the market easier in recent years as there's simply much less competition and more money buying no matter what the economic conditions or monetary policy are to exploit. I've written a blog post on my take in depth on the subject here: https://grizzlybulls.com/blog/time-in-the-market-vs-timing-t...

I view the Shannon algo going long dollar or long on a stock as a form of martingale, which is open for debate, but I was actually asking because I got slightly obsessed with disproving that it worked IRL, and started just messing with it, which got way past that and ended with me building my own model... and it seems to have interesting similarities to yours (superficially, at least). I've only recently put into practice. Purely because of my own risk tolerance, it's also long term, not intra-day, also not leveraged, and also only going long. One choice I made (for aesthetic reasons?) was to not use big data or anything beyond basic price histories, and to not use a neural net. I wanted something with simple, reproducible signal rules a human can understand, even at the expense of profit. I ran it through my own symbolic regression platform just to validate that I was onto something. I thought about posting it here and/or making it a paid service but it seemed to me (as someone said) that if anything like this makes enough money I wouldn't need to sell it, and would be better keeping it secret... plus I used to run a sports betting analysis site that was too much trouble and made me feel bad when predictions were off. As it is now, I'll only have to apologize to a couple of friends on a private email chain. Officially the target is 20% over S&P per year, but unofficially the results I'm seeing in the sim are 100% to 1500%.

I saw something it doesn't appear is present in your models, and I extrapolated from it into a program that chases that thing. Nonetheless it definitely seems like we've worked in a similar direction, as one potential application of my model is a simple long stock vs dollar trigger. And you've probably gleaned other insights I haven't. Offhand, would you be open to getting in touch and swapping strategies / pitfalls?

Very interesting, I'm glad to hear you are also building your own models and having success! I'm pretty tight-lipped about the inner-workings of mine, as unlike selling access to the generated signals, sharing details about the internals actually would hurt my edge and personal returns. That's why I've left the details pages intentionally high level and vague, in fact I think I may have erred on the side of too much detail already. I'm very impressed that you are seeing triple digit returns in the backtest of your model though without using leverage and long-only. Using long/short or modest leverage puts our platinum model into triple digits, but also increases the max drawdown and standard deviation of returns. I leave it up to the clients whether to use a riskier implementation such as this.

How long have you been running it live, and how far back does your backtest data run? How closely are you seeing real returns match backtest? After running live for over 2 years, I've seen up to 6 month periods where the model outperformed the backtest, but overall the models have underperformed to varying degrees as expected. Though I have spent hundreds of hours dedicated to removing noise and overfit bias, I've come to the conclusion it isn't possible to 100% remove it, which combined with expected market regime changes where legitimate patterns become unprofitable means that in the long run every model is essentially guaranteed to underperform its backtest, the question becomes simply to what degree. However, with a strong enough model, even underperforming the backtest can lead to substantial alpha which is what I and our members have experienced so far this year.

I've only been running it live for a couple months, so it's too soon to tell! Generally it wouldn't be expected to show any signals for the first 6 months or so. Additionally it's expected to weaken over time and need to be reset; 3 years seems about the optimal horizon.

The testing data I'm using goes back to Y2K. Overfitting is definitely a concern. I've proven that a delta that worked one year doesn't work the next, or what worked for one five-year span doesn't work on the next five. But what's interesting is that the local delta mean can be broadly broken into distinct eras by looking at sliding time windows. e.g. From 1998-2008, the best overall fit scales down well to each year. 2008-2012, 2012-2020, 2020-now... the last one is the great unknown, of course, and too early to call if I've gotten close to it. Any test that covers the covid disruption or the 2008 crisis has a nexus that skews future results to the high side, so I feel those can't be trusted. If I can beat the market by 20% in the first year, I'll be happy.

I'll keep an eye on your progress. Best of luck!

Reply on news.ycombinator.com