hi, thanks for the work on this. I am looking at markowitz_tuning.ipynb for the simulate() function.

def simulate(ht, ut):
    ret = torch.exp(logreturn1p_dist.sample())
    ht = ret * (ht + ut)                  <----- this line
    return ht, ret

from my understanding, ht should be (ret * (ht + ut))/(ret.T @ (ht + ut)) as per formula stated for how holdings evolve in 5.3 of the paper? This formula sums to 1 (as holdings should be), while the line in code doesn't.

Hi,

I would appreciate getting your help with the LQR example.
Following the LQR example, I notice that for a continuous problem, e.g., cart pole, with linearized dynamics at the fixed point, the LQR solution doesn't work. It returns "inf."

For example,

Consider the following dynamics matrices:
A = np.array([ [0, 0, 1, 0],
[0, -0.71707317, 0, 0],
[0, 0, 0, 1],
[0, 15.77560976, 0, 0]])

B = np.array([ [0],
[0],
[0.97560976],
[-1.46341463]])

The following code doesn't work:
So that you know, I changed the discrete ARE from your notebook to the continuous ARE.

###################################################
P = cp.Variable((n, n), PSD=True)
R0cvxpy = cp.Parameter((m, m), PSD=True)

objective = cp.trace(P@W)
constraints = [cp.bmat([
[R0cvxpy, B.T@P],
[P@B, Q0 + A.T@P + P@A]
]) >> 0, P >> 0]
R0cvxpy.value = R0
result = cp.Problem(cp.Maximize(objective), constraints).solve()
P_lqr = P.value
print(result)
##################################################

Returns: Inf

UserWarning: Solution may be inaccurate. Try another solver, adjusting the solver settings, or solve with verbose=True for more information.

I appreciate any help you can provide.

Best regards,
Solomon

hi all
@SteveDiamond @davidhallac @echu @bodono

im just wondering for the markovitz portfolio optimisation section of this paper.
how would you backtest it. would you retrain the model every time you re-balance.

I tried to follow the code but its a bit hard to tell how to use the weights in a backtest.

Kind regards,
Andrew