Not a good idea as taking the derivative of noisy measurements increases noise and therefore uncertainty.
So then, you take the correlation coefficient computed over the series and "apply" it to the last price of the dependent variable to predict the independent variable for the next period? Isn't this roughly equivalent to a simple linear regression where the dependent variable is lagged by one period?
The dependent var can only predict future values of itself. LinRegression(dependentvar,length) or Correlation(dependentvar[1],dependentvar[0],length) In theory the dependent var has a relationship with previous values of iv but in practice there isn't a relationship between dv and future values of iv. To predict iv requires LinRegression(iv,length) or Correlation(iv[1],iv[0],length); Yes the utility comes from modifying the correl default index so it calculates a leading rather than centered correlation. Which is a better predictor? LinRegression(c,length) or Correlation(iv[1],dv[0],length)? What about LinRegression(correl,length) or Correlation(correl[1],correl[0],length)?
You're quite correct on the dv/iv relationship, I mistakenly switched the terms in my previous post. I suspect the LinRegression would be the better predictor where you have a strong lead/lag relationship. But I'm not convinced that a lead/lag relationship exists in most markets.
How about a prediction of when it will end of why so many argue they don't like TA or don't like chart patterns. I bet the 95% who lose, most had charting up and or TA. The ones in the 5% group have figure out to use one, both, the other, neither or just argue they don't use TA when they do. What is that line "Most people have two opinions and neither one is right"? Lagging indicators keeps one from over trading, requires deeper retracements, keeps one longer in trends, less false signals. Many older traders like slower indicators, cause much time was spent studying price action besides indicators.
These aren't trick questions and I'm not an expert on anything. I've played with what's available. I can think of lots of things that are better than linear regression. JMA, EMA, HMA, SMA. I recall being mesmerized by straight line LR because due to it's nature it's the only obvious predictor, everything else is a curve. When viewing LR as a curve it's weaknesses are obvious. It fails. What I liked about it as a straight line was it includes SDev boundaries. At various times, with other curves than LR, I've used various methods for defining deviations from norm that I preferred over SD but I like SD best currently, fwiw, if that's enough disclosure to help someone they would of figured it anyway. No one can fake the drive or love necessary to succeed in this business. The default correl function in my platform is a simple binary signal. When I discovered Pearson coeff R I was mesmerized by it's mysteries I couldn't understand until a long time after taking everything apart. Not just pearson, every formula I could find. I've written my own correlation formulas, not so I could brag, simply because I needed to. It's really not hard to do nor anything to be bragging about. Study the binary correl for ideas. Forget about pearson and any form of LR, imo. If anyone ever begins to see how correl applied to preprocessed inputs, when combined with my simple but brilliant offset index ...is the core of all leading indicators ...then please stop arguing about lagging this and lagging that.
Hat tip to you! The textbook formulas in the public domain are best used for learning and reference. Then take the initiative to roll your own. That's how you make money in this business.