Yes, the engine failed, the car crashed (it was totaled) and there were no airbags. But airbags might not have been sufficient. Already, the composite signal of the In & Out strategy had for primary objective to protect from drawdowns. It was even the strategy's mantra: signal goes down, you get out of stocks; signal goes up, you move back in. For when the market was not moving up, you were in bonds or cash equivalents with low return expectations. Over last year, the signal (engine or whatever) was wrong 86% of the time. That makes the signal composition assumptions faulty and in need of repair, redesign, or at least reassessment.
I currently see the In & Out strategy as a classic case of curve fitting to the extreme. It had many modifications and tweaks while being developed on the Quantopian website. Its trading signals were fine-tuned over what should be considered its training and validation periods prior to October 2020. Based on portfolio metrics generated by those simulations prior to October 2020, it seemed to work fine, showed promise, including CAGR-wise.
This particular version was ported to QuantConnect in October 2020 and also had quite a few modifications. We could consider that from October 2020 to October 2021 (a whole year), the strategy was further refined and had its validation and walk-forward phase (a year is more than enough for that). Should we design these things to put them on walk-forward mode for years on end? At one point, those strategies have to go live or else be discarded.
In my previous post, I did not want to be too harsh which is why I used the term "walk forward" when describing the In & Out strategy. What I intended to use was "going live" after the strategy's one-year "walk forward" leading to October 2021. The strategy has been frozen in time since October 2021 and therefore any simulation up to October 2022 would be like if having traded this strategy live with its consequences (good or bad).
A simple simulation from October 2021 to October 2022 was sufficient to show that the strategy was clearly overfitted. It is the most reasonable explanation. From what should have been its live phase, the strategy went from an 86% win rate prior to October 2021 to a -86% loss rate from October 2021 to October 2022. This means that 86% of the trading signals were wrong or mistimed. Those were not sought-after portfolio metrics, especially if trading live.
Whatever this strategy was or has been up to October 2021, it does not matter so much for the person that traded that thing live from October 2021 to October 2022. It was simply a disaster, a real train wreck, a portfolio destroyer, providing no reprise.
What can be done now is a re-design, an overall restructuring. We could use the information that has been gained over those 14 years. But first, there is a need to demonstrate that the composite trading decision signal has value, not of the coincidental type once in a while but from a statistically significant point of view. However, there might not be sufficient data to make that case.
Here is the equation for this strategy, or any other, for that matter:
