WEBVTT

NOTE Can the model beat yesterday?: captions match the on-screen steps.

1
00:00:00.200 --> 00:00:05.712
Step 1 of 7. Eight audits re-check the 2022 results: three conclusions flip.

2
00:00:05.712 --> 00:00:14.579
Step 2 of 7. Five price models against 'tomorrow equals today', tested on later days: none wins.

3
00:00:14.579 --> 00:00:20.347
Step 3 of 7. The 2022 set-up shuffled the days: the Apple LSTM's 1.00% error looks respectable.

4
00:00:20.347 --> 00:00:24.996
Step 4 of 7. Train on the past, test on the future: 38.69% against 1.67% for yesterday's close.

5
00:00:24.996 --> 00:00:34.082
Step 5 of 7. Day by day: the network never saw prices this high, so it cannot extrapolate.

6
00:00:34.082 --> 00:00:42.446
Step 6 of 7. Diebold-Mariano on the same 690 days: gap +37.03 points (95% CI +33.06 to +41.00), p < 0.0001.

7
00:00:42.446 --> 00:00:51.680
Step 7 of 7. Five training seeds: every time-ordered run loses; on shuffled days only seed 42 wins.