You did not try to model accurately the 1st part of the trajectory and we have no handshakes between 17:10 and 18:30, ok with that, but the trajectories you used frankly cross the Indonesian airspace: shouldn't such trajectories have prompted a reaction from the Indonesian AF as they did in the past ?
I was trying to use the simplest possible model that will satisfy the data to the level of error in that data, which I estimate to be ~2Hz, one sigma. So I have used two parameters only, the speed of the first leg and the speed of the rest of the legs. As noted in section 7.9, the speeds of the first and second legs can be traded to move the position of the 18:29UT ping North, without affecting the fit. We are on very dodgy statistical ground here. There are only 4 reliable BFO data points at most, and I am fitting two parameters already. If I fit more parameters there is a danger of 'over-fitting', which means any data can be fitted. What the fitting process must avoid (of course the investigation knows this) is to avoid just joining the data dots with a complex model. Such a solution would definitely
not be correct, there is noise in the data and this has to addressed.
I have been concerned that presentations (to the families) have made attempts to bend the path round Indonesia. The ping rings in the slides shown were not those derived from the data log (or from the slide of satellite elevations). The precisely defined 18:29UT turning point in the slides seems to be an assumption about the navigation process of the flight - perhaps correct, but not supported by any fact of which I am aware.
Better in the BFO/BTO analysis to avoid using data that cannot be 100% verified. If the data is consistent with the radar data, a consistent track should appear as a statistical acceptable solution.
I think it is possible to reproduce the 3rd of the 3 handshakes at around 18:30: can you visualize the continous underlying time serie D1_aircraft+D2_aircraft ? (not simply its values at the handshake instants).
Yes, I think a course can be generated that will satisfy some of the data around ~18:29UT, but I am not sure it adds any weight to the fit, that is a degree of freedom has to be used up matching the added data point.
On the time data series I will work on that. It will require an assumption about how the course changes between ping-rings (as a function of time) - at the moment I have just modelled one course per leg. That can be done of course, but it is another assumption.