In fact these are two individual problems, which level should be used to optimize and which level should be used to calculate ECD.

If the non-covalent interactions cannot be neglected, then diffuse basis function must be used during optimization; however, as you mentioned, the SCF convergence will become much more difficult. There is a very nice way to bypass this problem, namely Grimme's gCP method(JCP, 136, 154101), which is way to correct BSSE problem without any additional cost, and it has been fully supported by ORCA. That means, you can safely use such as def2-SV(P) with gCP correction to optimize the systems containing evident non-covalent interactions, in this case SCF should not be difficult to converge and the computational cost is fairly low (especially if you use pure functionals, which can be substantially accelerated by resolution-of-identity (RI) technique in ORCA). A very very ideal level in ORCA to optimize large system containing non-covalent interactions is RI-BLYP/def2-SVP with gCP BSSE correction and DFT-D3 dispersion correction.

As for ECD calculation for large system containing several hundreds of atoms, 6-31G* or 6-311G* is good enough, diffuse functions are not needed. In my view, wB97XD is preferred over CAM-B3LYP. Worthnotingly, if you would like to reduce computational cost during ECD calculation, I highly suggest you pay attention to Grimme's sTDA method (see original paper JCP, 138,244104, as well as its extension to TDDFT and range-separated functional cases: Comp. Theor. Chem.,1040-1041,45 and PCCP,16,14408). By using this sTDA method, the total time cost of ECD is almost identical to single point calculation. sTDA program can be freely obtained from Grimme's webpage (http://www.thch.uni-bonn.de/tc/index.php?section=downloads&subsection=sTDA&lang=english).

Thank you, I'll probably try both approaches. As a side note, the structures are not optimized but taken from MD simulations -- so now I'm thinking it might be indeed a good idea to start with optimization to improve convergence, even if that's less rigorous from a thermodynamical point of view.

When you calculate ECD spectra by TD-DFT on snapshots taken from MD simulations, you don't have to optimize the selected structures before the calculations, otherwise you completely lose the information given by MD.

Apart from all the considerations on basis sets and functionals, which are indeed important, I guess that in your case 20 starting structures are not sufficient to obtain a proper statistical representation of your molecule at equilibrium, and consequently you are not getting good correlation with experimental values. You'll probably need to increase the number of snapshots from 20 to, say, 100 or 200 and, if necessary, simulate the system for longer times by MD.

Another thing you may need to consider, since you are already working with MD, is explicit solvation: PCM is a good approach, but it can fail if short-range solvent-solute interactions are influencing the chiroptical properties of your system. 

I appreciate your comment. I know 20 snapshots might be too few to get a *very good* estimate, but I'd expect at least several frames to give *reasonably good* energies, while in my case the maximum peak is systematically off by 30-70 nm and the minimum peak is roughly where it should be. Also, I am aware that geometry optimization is not the best idea for MD snapshots, but sometimes thermally distorted geometries are way harder to converge.

Explicit solvation is also something I considered, but even now I happen to run into memory troubles with my calculations, and here with additional solvent molecules things can only go worse. Anyway, if all else fails, I will sure try this approach with moderate basis sets.

Do you have hydroxyl groups close to cromophores in your molecule? If yes, then it may be that the electronic transition responsible for the shifted ECD band is very sensitive to the orientation of the hydroxyl groups. You can, for instance, test the position of the band when you force the hydroxyl groups to rotate.

If you see large shifts by rotating, then just a few explicit solvent molecules (water, if I understood correctly) around the hydroxyl groups may eventually solve your problem. I know that this can put you in some trouble with memory and SCF convergence, but hydroxyl groups can be very tricky when it comes to ECD.

Indeed, there are many hydroxyl groups in my molecule -- it's a polyene antibiotic with a polyol part; however, the orientations should be more or less random in my 20 frames (they are at least several ns apart), and I do not see that much variation in the band positions. But again, thanks for this remark -- I'll sure bear it in mind in case Grimme's sTDA fails to significantly improve the results.