Nearest Neighbor Extrapolation of Price [Loxx]I wasn't going to post this because I don't like how this calculates by puling in the Open price, but I'm posting it anyway. This does work in it's current form but there is a. better way to do this. I'll revisit this in the future.
Anyway...
The k-Nearest Neighbor algorithm (k-NN) searches for k past patterns (neighbors) that are most similar to the current pattern and computes the future prices based on weighted voting of those neighbors. This indicator finds only one nearest neighbor. So, in essence, it is a 1-NN algorithm. It uses the Pearson correlation coefficient between the current pattern and all past patterns as the measure of distance between them. Also, this version of the nearest neighbor indicator gives larger weights to most recent prices while searching for the closest pattern in the past. It uses a weighted correlation coefficient, whose weight decays linearly from newer to older prices within a price pattern.
This indicator also includes an error window that shows whether the calculation is valid. If it's green and says "Passed", then the calculation is valid, otherwise it'll show a red background and and error message.
Inputs
Npast - number of past bars in a pattern;
Nfut -number of future bars in a pattern (must be < Npast).
lastbar - How many bars back to start forecast? Useful to show past prediction accuracy
barsbark - This prevents Pine from trying to calculate on all past bars
Related indicators
Hodrick-Prescott Extrapolation of Price
Itakura-Saito Autoregressive Extrapolation of Price
Helme-Nikias Weighted Burg AR-SE Extra. of Price
Weighted Burg AR Spectral Estimate Extrapolation of Price
Levinson-Durbin Autocorrelation Extrapolation of Price
Fourier Extrapolator of Price w/ Projection Forecast
Extrapolator
Hodrick-Prescott Extrapolation of Price [Loxx]Hodrick-Prescott Extrapolation of Price is a Hodrick-Prescott filter used to extrapolate price.
The distinctive feature of the Hodrick-Prescott filter is that it does not delay. It is calculated by minimizing the objective function.
F = Sum((y(i) - x(i))^2,i=0..n-1) + lambda*Sum((y(i+1)+y(i-1)-2*y(i))^2,i=1..n-2)
where x() - prices, y() - filter values.
If the Hodrick-Prescott filter sees the future, then what future values does it suggest? To answer this question, we should find the digital low-frequency filter with the frequency parameter similar to the Hodrick-Prescott filter's one but with the values calculated directly using the past values of the "twin filter" itself, i.e.
y(i) = Sum(a(k)*x(i-k),k=0..nx-1) - FIR filter
or
y(i) = Sum(a(k)*x(i-k),k=0..nx-1) + Sum(b(k)*y(i-k),k=1..ny) - IIR filter
It is better to select the "twin filter" having the frequency-independent delay Тdel (constant group delay). IIR filters are not suitable. For FIR filters, the condition for a frequency-independent delay is as follows:
a(i) = +/-a(nx-1-i), i = 0..nx-1
The simplest FIR filter with constant delay is Simple Moving Average (SMA):
y(i) = Sum(x(i-k),k=0..nx-1)/nx
In case nx is an odd number, Тdel = (nx-1)/2. If we shift the values of SMA filter to the past by the amount of bars equal to Тdel, SMA values coincide with the Hodrick-Prescott filter ones. The exact math cannot be achieved due to the significant differences in the frequency parameters of the two filters.
To achieve the closest match between the filter values, I recommend their channel widths to be similar (for example, -6dB). The Hodrick-Prescott filter's channel width of -6dB is calculated as follows:
wc = 2*arcsin(0.5/lambda^0.25).
The channel width of -6dB for the SMA filter is calculated by numerical computing via the following equation:
|H(w)| = sin(nx*wc/2)/sin(wc/2)/nx = 0.5
Prediction algorithms:
The indicator features the two prediction methods:
Metod 1:
1. Set SMA length to 3 and shift it to the past by 1 bar. With such a length, the shifted SMA does not exist only for the last bar (Bar = 0), since it needs the value of the next future price Close(-1).
2. Calculate SMA filer's channel width. Equal it to the Hodrick-Prescott filter's one. Find lambda.
3. Calculate Hodrick-Prescott filter value at the last bar HP(0) and assume that SMA(0) with unknown Close(-1) gives the same value.
4. Find Close(-1) = 3*HP(0) - Close(0) - Close(1)
5. Increase the length of SMA to 5. Repeat all calculations and find Close(-2) = 5*HP(0) - Close(-1) - Close(0) - Close(1) - Close(2). Continue till the specified amount of future FutBars prices is calculated.
Method 2:
1. Set SMA length equal to 2*FutBars+1 and shift SMA to the past by FutBars
2. Calculate SMA filer's channel width. Equal it to the Hodrick-Prescott filter's one. Find lambda.
3. Calculate Hodrick-Prescott filter values at the last FutBars and assume that SMA behaves similarly when new prices appear.
4. Find Close(-1) = (2*FutBars+1)*HP(FutBars-1) - Sum(Close(i),i=0..2*FutBars-1), Close(-2) = (2*FutBars+1)*HP(FutBars-2) - Sum(Close(i),i=-1..2*FutBars-2), etc.
The indicator features the following inputs:
Method - prediction method
Last Bar - number of the last bar to check predictions on the existing prices (LastBar >= 0)
Past Bars - amount of previous bars the Hodrick-Prescott filter is calculated for (the more, the better, or at least PastBars>2*FutBars)
Future Bars - amount of predicted future values
The second method is more accurate but often has large spikes of the first predicted price. For our purposes here, this price has been filtered from being displayed in the chart. This is why method two starts its prediction 2 bars later than method 1. The described prediction method can be improved by searching for the FIR filter with the frequency parameter closer to the Hodrick-Prescott filter. For example, you may try Hanning, Blackman, Kaiser, and other filters with constant delay instead of SMA.
Related indicators
Itakura-Saito Autoregressive Extrapolation of Price
Helme-Nikias Weighted Burg AR-SE Extra. of Price
Weighted Burg AR Spectral Estimate Extrapolation of Price
Levinson-Durbin Autocorrelation Extrapolation of Price
Fourier Extrapolator of Price w/ Projection Forecast
Modified Covariance Autoregressive Estimator of Price [Loxx]What is the Modified Covariance AR Estimator?
The Modified Covariance AR Estimator uses the modified covariance method to fit an autoregressive (AR) model to the input data. This method minimizes the forward and backward prediction errors in the least squares sense. The input is a frame of consecutive time samples, which is assumed to be the output of an AR system driven by white noise. The block computes the normalized estimate of the AR system parameters, A(z), independently for each successive input.
Characteristics of Modified Covariance AR Estimator
Minimizes the forward prediction error in the least squares sense
Minimizes the forward and backward prediction errors in the least squares sense
High resolution for short data records
Able to extract frequencies from data consisting of p or more pure sinusoids
Does not suffer spectral line-splitting
May produce unstable models
Peak locations slightly dependent on initial phase
Minor frequency bias for estimates of sinusoids in noise
Order must be less than or equal to 2/3 the input frame size
Purpose
This indicator calculates a prediction of price. This will NOT work on all tickers. To see whether this works on a ticker for the settings you have chosen, you must check the label message on the lower right of the chart. The label will show either a pass or fail. If it passes, then it's green, if it fails, it's red. The reason for this is because the Modified Covariance method produce unstable models
H(z)= G / A(z) = G / (1+. a(2)z −1 +…+a(p+1)z)
You specify the order, "ip", of the all-pole model in the Estimation order parameter. To guarantee a valid output, you must set the Estimation order parameter to be less than or equal to two thirds the input vector length.
The output port labeled "a" outputs the normalized estimate of the AR model coefficients in descending powers of z.
The implementation of the Modified Covariance AR Estimator in this indicator is the fast algorithm for the solution of the modified covariance least squares normal equations.
Inputs
x - Array of complex data samples X(1) through X(N)
ip - Order of linear prediction model (integer)
Notable local variables
v - Real linear prediction variance at order IP
Outputs
a - Array of complex linear prediction coefficients
stop - value at time of exit, with error message
false - for normal exit (no numerical ill-conditioning)
true - if v is not a positive value
true - if delta and gamma do not lie in the range 0 to 1
true - if v is not a positive value
true - if delta and gamma do not lie in the range 0 to 1
errormessage - an error message based on "stop" parameter; this message will be displayed in the lower righthand corner of the chart. If you see a green "passed" then the analysis is valid, otherwise the test failed.
Indicator inputs
LastBar = bars backward from current bar to test estimate reliability
PastBars = how many bars are we going to analyze
LPOrder = Order of Linear Prediction, and for Modified Covariance AR method, this must be less than or equal to 2/3 the input frame size, so this number has a max value of 0.67
FutBars = how many bars you'd like to show in the future. This algorithm will either accept or reject your value input here and then project forward
Further reading
Spectrum Analysis-A Modern Perspective 1380 PROCEEDINGS OF THE IEEE, VOL. 69, NO. 11, NOVEMBER 1981
Related indicators
Levinson-Durbin Autocorrelation Extrapolation of Price
Weighted Burg AR Spectral Estimate Extrapolation of Price
Helme-Nikias Weighted Burg AR-SE Extra. of Price
Itakura-Saito Autoregressive Extrapolation of Price
Modified Covariance Autoregressive Estimator of Price
