WLS eviews options for weights

[Patric9871]

Hi,
I am modelling financial data with very high frequency. I have more than 150 000 observations (1 minute rates of return), but many of them are equal to 0, because there is not always transaction made every minute.
I am using Eviews 9, 64-bit ver

My model looks like that:
rate(t) = f(rate(t-1), rate(t-2), ... rate(t-k), dummy1(t-1), dummy1(t-2), ..., dummy1(t-j), dummy2(t-1), dummy2(t-2), ..., dummy2(t-j)) + e(t) [1]
In order to compare coefficients from OLS with HAC and two-step weighted least squares (2WLS), method that should eliminate ARCH problem I am estimated the model via WLS in the following manner:
a) estimation of standard OLS regression [1]
b) save the resids from OLS from point a)
c) save the absolute value of resids from OLS (abs_resid)
d) run auxiliary regression:
abs_resid (t) = abs_resid(-15 to-1), dummy1(t-1), dummy1(t-2), ..., dummy1(t-j), dummy2(t-1), dummy2(t-2), ..., dummy2(t-j) + u(t) [2]
e) save fitted values from regression in point d)
f) save weights as: 1/(fitted values ^2)
g) run weighted least squares

Using eviews help:
"The Type dropdown is used to specify the form in which the weight data are provided. If, for example, your weight series VARWGT contains values proportional to the conditional variance, you should select Variance. Alternately, if your series INVARWGT contains the values proportional to the inverse of the standard deviation of the residuals you should choose Inverse std. dev. "

My doubts:
1. Here I am not sure which option I should choose? What should be proper option for above type of model?
2. Besides that I receive error “Near singular matrix error. Regressors may be perfectly collinear” when I choose “inverse standard deviation” or “inverse variance”. Standard OLS does not create such error, so how it can be that weighting does it?

Edit: 1/4/19
Any help?
3. What would be your strategy to eliminate autocorrelation after running WLS model? WLS generally deals with heteroskedasticty but not with autocorrelation, and we cannot include ar() ma() in WLS in eviews …
4. I tried to use ARCH models but autocorrelation in such models remains very persistent (see below):
Autocorrelation Partial Correlation AC PAC Q-Stat Prob*

| | | | 2 0.021 0.021 23.326 0.000
| | | | 3 0.026 0.025 57.812 0.000
| | | | 4 0.048 0.048 181.85 0.000
*| | *| | 5 -0.109 -0.111 811.16 0.000
| | | | 6 0.025 0.024 844.16 0.000
| | | | 7 0.036 0.039 913.93 0.000
| | | | 8 0.027 0.030 953.67 0.000
| | |* | 9 0.068 0.077 1198.3 0.000
*| | *| | 10 -0.165 -0.188 2628.7 0.000
| | | | 11 0.044 0.048 2730.1 0.000
| | | | 12 0.051 0.062 2865.7 0.000
| | | | 13 0.041 0.047 2954.7 0.000
| | | | 14 -0.007 0.019 2957.3 0.000
| | *| | 15 -0.001 -0.067 2957.4 0.000
| | | | 16 -0.004 0.004 2958.3 0.000
| | | | 17 -0.000 0.020 2958.3 0.000
| | | | 18 -0.004 0.008 2959.3 0.000

[Patric9871]

Any help, please?