I combined two probability distribution functions from broad-band fitting (PIPE2D-539) and chi-square fitting of spectra (PIPE2D-835) to find the best fit model template of a flux standard. The combination of two PDFs is a one-to-one merge.
 
The following figures are the results of the simulated spectra with ETC and the HSC 5-bands photometry with S/N=100. The input stellar parameters of the ETC simulation are Teff = 7000 K, log(g/(cm/s^2)) = 4.5, Z = 0, and [a/Fe] = 0.0.
 
A combined PDF of an r = 13 AB mag case, the PDF of the broad-band fitting, and the spectral fitting, from the top, respectively.

The input spectrum (ETC) and the best-fit spectrum. 

 
A ratio of the best-fit spectrum to the input. There is a large deviation at a short wavelength of < 430 nm. This could be from small differences of Teff and log(g).

These figures are the case of r = 19 mag.
The simulated spectrum is noisier, and therefore the constraint from spectral fitting is weak. As a result, the combined PDF is dominated by the broad-band fitting.

A combined PDF.

Broad-band fitting PDF.

Spectral fitting PDF.

The best-fit spectrum and the input spectrum.

The flux ratio of the best-fit to the input.

This is the fitting accuracy as a function of the magnitude. The y-axis is the median absolute deviation of the flux ratio of the best-fit to the input.

The measured dispersions in the flux ratio of the best-fit spectrum to the simulation spectrum are dominated by the noise in the simulation spectra. I add the dispersion only from the noise in a spectrum in the above figure (open red circles). The measured dispersions are close to the noise dispersions.

In the three arms, the blue arm has an offset from the noise dispersions. This is due to a mismatch of the best-fit model to the simulation in the blue arm. We should improve this. 

 
The flux ratios of the best-fit and the simulation are small compared with the noise in the spectra in this test case. We should check other cases to see the fitting goodness.

The above test shows that the fitting accuracy is about 10% for 19 magnitude stars. Although this is higher than the 1% accuracy we are aiming at, we think we can improve it by the following approaches. 

1. Because we will have several tens of standard stars in an FoV, we can statistically reduce the accuracy by combining all the standards. 

2. The dispersion of the flux ratio is dominated by the noise in the simulated spectrum. Because we expect that a flux calibration vector is smooth along the wavelength, we can fit it with a function and reduce the dispersions. 

These works to improve the fitting accuracy will be addressed during the implementation phase. 

You can't assume that the flux calibration vector is completely smooth in wavelength: it should include the atmospheric absorption bands.

Yes. Please let me make a few corrections and add some words. The flux calibration vector includes both a smooth component and a non-smooth one.

Although this idea needs further investigation and discussion, for example, we will fit the smooth component in the flux calibration vector with a function after masking the atmospheric absorption bands to separate the smooth component from the vector. We will use the averaged flux calibration vector with a high S/N over an FoV for this. The non-smooth component made up of the atmospheric absorption bands will be reproduced using a template or a model. By combining the fitted smooth component and the reproduced non-smooth one, we will get a more accurate flux calibration vector.