Le Gardien · the record

What we said, and what happened

Every account that came within 5% of its liquidation price gets flagged by the same engine that protects clients. This page publishes what became of them — with no addresses, ever.

19.4%of flagged accounts were liquidated within 24 hours

10.7%of everyone else — the control group

1.8×more likely to die once flagged

Over 29 days we flagged 59 799 accounts. 45 056 of them could be checked against the exchange's own liquidation records, and 8724 were liquidated within the next 24 hours.

Day by day

Day	Flagged	Checked	Liquidated	Rate
18 Oct	3369	2324	391	17%
17 Oct	3166	2163	276	13%
16 Oct	3400	2238	321	14%
15 Oct	3429	2192	528	24%
14 Oct	1594	1282	339	26%
13 Oct	1544	1205	125	10%
12 Oct	1714	1401	441	31%
11 Oct	1574	1289	150	12%
10 Oct	1542	1156	57	5%
09 Oct	1619	1212	102	8%
08 Oct	1910	1530	617	40%
07 Oct	1899	1502	517	34%
06 Oct	1775	1428	138	10%
05 Oct	1804	1404	123	9%
04 Oct	1726	1281	128	10%
03 Oct	1687	1250	63	5%
02 Oct	1952	1516	457	30%
01 Oct	1850	1444	220	15%
30 Sep	1888	1449	198	14%
29 Sep	1951	1486	277	19%
28 Sep	2029	1553	428	28%
27 Sep	1899	1477	206	14%
26 Sep	1870	1426	131	9%
25 Sep	1996	1496	309	21%
24 Sep	2012	1537	356	23%
23 Sep	2202	1740	567	33%
22 Sep	2102	1674	330	20%
21 Sep	2236	1775	557	31%
20 Sep	2060	1626	372	23%

How to break this number

Three rules keep it honest, and each one makes the figure smaller:

We only count what we actually checked. Liquidations are read account by account from the exchange's own fill records. An account we flagged but never read afterwards is excluded, and the count of excluded ones is the gap between "flagged" and "checked" above. Without that rule we would be picking our evidence after the fact.

The last twenty minutes do not count. A flag raised minutes before a liquidation is not a prediction — it is the same event seen twice. Those are dropped.

The control group is drawn the same way. Accounts we never flagged are read from the same random sample, over the same days. Otherwise we would be measuring our own attention, not our accuracy.

A flagged account that survives is not a failure: often the person reacted, or the market turned. This is not a score out of ten. It is a measure of how much a flag means, next to the base rate.

And your own account? The same engine reads any address on request: open its file — every position the exchange ever force-closed on it, what it cost, and where it stands today. Read-only, no signature, nothing stored.