Thanks for the detailed response, I'll do my best to answer:
1. Internally, the model does work on a sort of weightings / probability system but it's not so simple as boiling down to a single probability figure, i.e. if it hits over 60% we trigger buy or less than 40% trigger sell. Instead there are a few phases in signal change identification starting with an attempt to classify overall fundamentals and market move and moving down to the more granular indicators based on each level of abstraction. Certain low level indicators mean very different things based on the higher level classifier output. However, I could feasibly attempt to boil down to some sort of confidence or probability figure and expose it. I'll need to do some rigorous internal testing to ensure that it conveys accurate information first, but it's a good suggestion.
2. I think the fairest way to report the results is with a 1x/0x (long with no leverage on buy signal, market neutral on short) and then let the users decide whether to use leverage or to short the market during sell signals. Personally I have used leverage here and there and I tend to dial it up or down based on how well the model seems to be performing at a given time. However, I think it's important to note that no model can ever be 100% free from overfitting, and while a great deal of effort has been made to generalize and widen params and heuristics, there's bound to still be some and even I don't know to what degree. This, plus the fact that algotrading becomes more competitive every year means that I think it's prudent to conservatively expect returns somewhere in the 50-100% wide range of the past, which could also mean drawdowns a bit larger than the backtest shows. And of course, black swan risk is ever present. Finally, the models only go back to 2009, so we don't have great data to indicate how they might perform in a more harsh bear market such as 2007-2009 or 2000-2003. This is an unfortunate coincidence as one of the key indicators, Vix Futures, only came to be in late 2004 and the earliest I could find intraday data broken into distinct contracts (not just continuous front month) was early 2009.
As far as my occasional contradictory past reporting of my own financial position. There's a couple reasons this could be--at times I've included less liquid assets (private company shares, family business interest, 401k, home equity, etc), and at others I've only reported what's immediately available to me and highly liquid (bot account + emergency fund, etc). I feel I've probably erred on the side of sharing too much personal financial details over the years, but it is what it is. Secondly, I've not always been a stable investor. I hate to admit that back in the day I was a bit of a WSBer at times. Never consistently, but sometimes I'd get caught up in the FOMO and make a rash (and often too large) bet on an earnings report or something and (though I won a few times) usually end up losing my ass. As such my net worth was pretty volatile between 2013-2018. I've since given that all up and the bot has helped assuage my inner gambler. I also used to follow some other folks' timing systems back then and learned the hard way that most don't work, so all skepticism is warranted. Even I like to keep a skeptical mind about the future of my own models which is why I reported the wide range of expected outcome above
1. Internally, the model does work on a sort of weightings / probability system but it's not so simple as boiling down to a single probability figure, i.e. if it hits over 60% we trigger buy or less than 40% trigger sell. Instead there are a few phases in signal change identification starting with an attempt to classify overall fundamentals and market move and moving down to the more granular indicators based on each level of abstraction. Certain low level indicators mean very different things based on the higher level classifier output. However, I could feasibly attempt to boil down to some sort of confidence or probability figure and expose it. I'll need to do some rigorous internal testing to ensure that it conveys accurate information first, but it's a good suggestion.
2. I think the fairest way to report the results is with a 1x/0x (long with no leverage on buy signal, market neutral on short) and then let the users decide whether to use leverage or to short the market during sell signals. Personally I have used leverage here and there and I tend to dial it up or down based on how well the model seems to be performing at a given time. However, I think it's important to note that no model can ever be 100% free from overfitting, and while a great deal of effort has been made to generalize and widen params and heuristics, there's bound to still be some and even I don't know to what degree. This, plus the fact that algotrading becomes more competitive every year means that I think it's prudent to conservatively expect returns somewhere in the 50-100% wide range of the past, which could also mean drawdowns a bit larger than the backtest shows. And of course, black swan risk is ever present. Finally, the models only go back to 2009, so we don't have great data to indicate how they might perform in a more harsh bear market such as 2007-2009 or 2000-2003. This is an unfortunate coincidence as one of the key indicators, Vix Futures, only came to be in late 2004 and the earliest I could find intraday data broken into distinct contracts (not just continuous front month) was early 2009.
As far as my occasional contradictory past reporting of my own financial position. There's a couple reasons this could be--at times I've included less liquid assets (private company shares, family business interest, 401k, home equity, etc), and at others I've only reported what's immediately available to me and highly liquid (bot account + emergency fund, etc). I feel I've probably erred on the side of sharing too much personal financial details over the years, but it is what it is. Secondly, I've not always been a stable investor. I hate to admit that back in the day I was a bit of a WSBer at times. Never consistently, but sometimes I'd get caught up in the FOMO and make a rash (and often too large) bet on an earnings report or something and (though I won a few times) usually end up losing my ass. As such my net worth was pretty volatile between 2013-2018. I've since given that all up and the bot has helped assuage my inner gambler. I also used to follow some other folks' timing systems back then and learned the hard way that most don't work, so all skepticism is warranted. Even I like to keep a skeptical mind about the future of my own models which is why I reported the wide range of expected outcome above