The zombie feed: how a delisted coin fabricated 314 winning paper trades
A frozen perp at 0.37621 against a live coin near 0.08 booked 314 fake paper wins. Purging them moved recorded P&L from $6,350 to $5,287.
This was paper trading. No live order was placed, and no live capital was at risk. A reader who thinks real money moved will misread every number that follows.
What did move is a research ledger. Between March and June 2026, a paper book recorded 314 winning trades on a coin that was no longer trading. Each one was a short opened at 0.376 against a market near 0.08. Each one closed as a take-profit. The win rate on those rows was 100 percent, because none of them were real. Purging them moved recorded paper P&L from $6,350 to $5,287.
That is not a curiosity. Fabricated wins in a research ledger are how a strategy gets promoted on evidence that never existed. This post is the incident, the two feeds that disagreed, and the sanity guard that should have caught the disagreement and did not.
Research write-up, not financial advice.
How it showed up
A paper book that trades many names can hide a bad feed inside ordinary variance. This one did not. One ticker was a perfect record: 314 closed rows, every one a take-profit, entries clustered on 0.37621. That is not a hot streak. It is a fingerprint. A live coin near 0.08 does not fill 314 shorts at 0.376. Once you look at the entry price instead of the win rate, the rest of the incident is mechanical.
Two prices, one ticker
In September 2024, MATIC rebranded to POL. On Coinbase, MATIC-USD is a delisted product. The live coin, POL, traded near 0.08.
Hyperliquid kept the old perpetual market alive. Its mid feed still returned MATIC at 0.37621, frozen at the last-traded price from early 2025. A zombie: a ticker that looks live, quotes a number, and does not move.
The paper trader took that mid as the entry. The candles it marked the position against were not the same coin. The OHLCV fetcher maps the old MATIC pair onto the successor, so barriers and P&L resolved against POL, near 0.08.
A short opened at 0.376 against a market at 0.08 is immediately about 79 percent in the money. The paper book treated the frozen mid as a fill, then as a take-profit, then as a win. It did this 314 times, from 2026-03-15 until the fix shipped. Every one of those rows took profit. None of them lost.
The guard that was almost right
There was already a sanity check on entry price. It compared the mid used for the fill against the candle market, and it rejected the entry if those two disagreed by more than 5x. Denomination bugs and stale feeds were exactly the class of failure it was written for.
0.37621 divided by 0.08 is about 4.7x. The bound was 5x. The discrepancy walked under the tripwire.
That is the detail that makes this incident worth writing down. The guard was the right shape: two independent prices, a maximum ratio, a hard reject. It was wrong by a hair. A missing check is a boring story. A check that was present, reasonable, and slightly too loose is the one that keeps happening to careful systems.
We tightened the bound from 5x to 2x, and we dropped the delisted ticker from the active set. The successor was not silently swapped in. The tighter bound now rejects discrepancies above 2x. It does not prove that a price below that threshold is fresh. Removing the delisted ticker is what closed this MATIC path. The bound will skip some genuine listing-day spikes on thin names. That is the direction the miss told us to err.
What the ledger actually said
On the closed paper book, before the purge:
| Slice | Paper P&L |
|---|---|
| Recorded closed book | $6,350 |
| 314 fabricated zombie shorts | +$1,063 |
| After purge | $5,287 |
Fabrication was 16.7 percent of recorded paper P&L. We backed up the 314 rows, then deleted them. Twelve legitimate successor-coin rows stayed in the book. The corrected closed ledger is $5,287. Removing a fabricated slice from a ledger is bookkeeping. It is not a performance result, and it is not a validated edge.
MATIC was the sole remaining fabrication source in that book. We are not going to publish the rest of the table. This post is about the rows that should never have been there.
What we are not claiming
- This was paper only. The paper backend booked fills from a frozen mid without placing a real order.
- The remaining paper P&L is not an edge. A cleaner ledger is not a better strategy. It is a ledger that stopped lying.
Never trust a single venue’s mid for a ticker that might be delisted. Cross-check it against an independent feed. And when you put a ratio bound on that check, remember that the discrepancy you actually get will sit just inside a round number.
Paper trading only. No live capital was at risk. Research write-up, not financial advice.
The guard was the right shape. It was wrong by a hair.
- +A sanity bound that is the right shape can still miss by sitting just above the discrepancy
- +Paper fills that never place a real order can still poison a research ledger
- +Never trust a single venue mid for a ticker that might be delisted: cross-check an independent feed
- +Fabricated wins are how a strategy gets promoted on evidence that never existed
- ×Hyperliquid kept a delisted perp alive at a frozen mid of 0.37621
- ×Barriers and P&L resolved against successor-coin candles near 0.08
- ×The entry-price guard's 5x bound let a 4.7x zombie through
- ×314 paper rows booked as 100 percent winners before anyone noticed
Our LLM Polymarket bot was just reading the market price back to us
Aggregate Brier score 0.12. And it lost to the market in every single edge bucket. A postmortem of an LLM forecaster that turned out to be an expensive mirror.
The look-ahead bug that made our backtest look worse
Our Chandelier exit peeked one bar ahead, cutting winners short instead of inflating the result. A fixed backtest is not a found edge: ETH 4h went from -16.2% to +185.0%.
무엇이 돌아갔고, 무엇이 출시됐고, 무엇이 죽었는지 - 숫자와 함께. 스레드도, 과장도 없습니다.