Get ready for lift off, here comes the Gradient Adaptive Dual-Fueled RSI. As promised since before the New Year, this post pulls together concepts from several prior articles exploring how one might develop a superior adaptive indicator on one level, then improving that base indicator’s performance again by adaptively observing and responding to the system’s own equity curve.
The Basic Systems
The base system employs a MACD-like differential between two Relative Strength Indicators (“RSI“) of varying lengths. The first RSI is of a very short periodicity that is adaptive according to current volatility and efficiency of movement characteristics. The second is fixed at over a week in length (actually an average of two to avoid the risk of selecting a “magic” solution), providing a relative pivot to better capture price wavelets. One can imagine doing this again over an even longer period to actualize the concept suggested by David Varadi in his recent post, “Detrending Oscillators.”
The difference between the two RSIs is then normalized by dividing by near-term volatility, with the resulting calculation percent ranked using learning lengths that vary again according to volatility measures with more stable, low volatility periods warranting greater look back periods to reduce signal sensitivity.
The resulting frictionless equity curve trading the SPY, a liquid S&P500 ETF proxy, looks like so:
Not a bad start and a decent improvement over its predecessor approach found here, “Relative Volatility & Comparative Mean Reversion Performance,” but we can do better yet. The header to this section is entitled “Basic Systems,” plural…. The second is just buy and hold! As discussed in this “Seeking Linearity” piece, round-trip trading costs can be reduced significantly and marginal short-side trades avoided altogether when volatility is falling and prices are up-trending as defined by long-term moving averages.
Adaptive System Normalization & Regime Switching
The described fully normalized regime adaptive variant is shown below in blue (“Adaptive Unlevered”). Notice, in particular, the enhanced performance as compared to the basic variant above during serial bull-market environments:
So let’s recap features of the indicator at this point as compared to those discussed in this follow-on “Seeking Linearity” article:
- Daily Return Smoothing/ Noise Reduction
- Adaptive Cycle Consideration (Improved)
- Nonparametric Volatility Normalization
- Rolling Real-World Distribution Factoring (Improved)
- Adaptive Environmental Rule Shifting (Improved)
While I’m not promoting the Adaptive Dual-RSI as the end-all-be-all of indicators, I think you will agree that it hits on all the points above, and that they are critical considerations for any modern indicator’s construction.
Gradient & Expectation Exposure Filter
The last element focuses on variable exposure according to base system performance, as alluded to in the later “Seeking Linearity” post. Because the base indicator is so consistent, the overlay shown in red (“Gradient Boosted”) effectively leverages exposure higher in 25% increments up to 2:1 as the system progressively under performs, assuming mean reversion by the system itself. Re-read that last sentence. Yes, systems themselves maybe mean-reverting over varying time frames.
This approach could obviously prove catastrophic if the system serially spun out of control. I therefore employed a basic equity curve expectation minder to automatically turn the system off in the event of failure, as nicely described here by Dave at “MindMoneyMarkets.” Interestingly, however, the gradient exposure increases the win-rate of the binary base signals by nearly 10% points.
Over the 2,350 days tested here, the system provided at 74% hypothetical winning rate of trade, with losing trades slightly larger than winners by a factor of 1.2, and an average directional trade signal lasting approximately three to four days. Net exposure over the entirety of the period was between +20% and +30% long, including leveraged periods. Obviously slippage would take a toll on such a model, but I nevertheless believe the core concepts presented here are worthy of further research.
I’ll also be first to grant that the result looks curve fit and I’d never suggest trading a single system, but really its just a simple layering of applied core concepts at a variety of levels — lift off indeed! I hope you have found some of the ideas worth investigating on your own from this series.
Never Investment Advice
Related posts:
