| 00:00 - 00:03 | The covariance matrix |
| 00:04 - 00:05 | becomes the quadratic term |
| 00:05 - 00:07 | and it is guaranteed to be PSD. |
| 00:08 - 00:12 | So we can input it as a Q in a convex QP |
| 00:12 - 00:15 | and gradient descent should do the job. |
| 00:17 - 00:19 | But what you really want... |
| 00:19 - 00:21 | ... is a conic form with the Cholesky factor... |
| 00:24 - 00:26 | Sir... |
| 00:27 - 00:28 | We don't know... |
| 00:31 - 00:33 | We couldn't make it stable enough ... |
| 00:34 - 00:36 | we tried but there are always negative eigenvalues! |
| 00:53 - 00:58 | Markowitz, Cholesky, Numpy, stay in the room... |
| 01:13 - 01:15 | What do you mean? |
| 01:15 - 01:17 | Have I not made it clear!?? |
| 01:18 - 01:23 | Factorize your damn covariance matrix |
| 01:25 - 01:28 | and input it in conic form! |
| 01:29 - 01:31 | How come you don't get it? |
| 01:31 - 01:34 | And then you avoid the quadratic Q matrix |
| 01:34 - 01:37 | with those horrible singularities |
| 01:37 - 01:40 | that are responsible for huge violations! |
| 01:40 - 01:42 | Sir, we could not do the Cholesky... |
| 01:42 - 01:46 | Because you have not tried enough!!! |
| 01:46 - 01:48 | But Sir, the numerical issues... |
| 01:48 - 01:52 | Would not be a problem if your data wasn't rubbish! |
| 01:53 - 01:54 | Crap in, crap out, remember!?? |
| 01:56 - 01:57 | Scale your data! |
| 01:57 - 02:00 | Remove those huge coefficients! |
| 02:00 - 02:03 | Get the objective norm close to 1! |
| 02:04 - 02:08 | Replace the norm square with the norm! |
| 02:08 - 02:13 | And don't tell me it is impossible to scale the inputs |
| 02:14 - 02:16 | because it never ever is! |
| 02:17 - 02:21 | Unless your problem is illposed as hell! |
| 02:27 - 02:29 | For so many years |
| 02:30 - 02:34 | we have I advocated the conic form |
| 02:34 - 02:36 | and they still don't get it. |
| 02:41 - 02:42 | We even made FUSION |
| 02:43 - 02:47 | explicitly without quadratic terms |
| 02:48 - 02:53 | so they eventually get the point. |
| 02:54 - 02:56 | But no, who'd have thought |
| 02:56 - 02:59 | that people still compute X^T*X |
| 03:00 - 03:02 | just to make us factorize it back again!! |
| 03:04 - 03:07 | Don't worry. You never understood Sharpe ratio anyway. |
| 03:14 - 03:16 | What about the factor model? |
| 03:19 - 03:23 | It is... perfect in these circumstances. |
| 03:25 - 03:26 | Fits so well. |
| 03:31 - 03:33 | The factor model trick |
| 03:40 - 03:46 | is that risk = D + F^T * G * F with a small G easy to factorize. |
| 03:46 - 03:49 | And then... |
| 03:53 - 03:56 | ...the conic model is so much smaller. |
