NewByPerp · Bybit USDT perpetuals · data to 2026-09-28

Shorting new Bybit listings after the pump

A condition-based short on freshly listed crypto perps looks strong in-sample. Once the parameter search is tested honestly on unseen months, the edge mostly disappears.

Out-of-sample return per trade, config re-selected monthly
+0.7%
33 trades · profit factor 1.05 · p = 0.14 vs random shorts (not significant)
+14.8%
In-sample per trade (22 trades)
+7.8%
Headline config on the same months (hindsight)
−4.1%
Random shorts, same months
1.8 / mo
Signal frequency
Where the edge goes
Average net return per trade (after fees, slippage and funding), from the fully in-sample number down to an honest out-of-sample test.

Bottom line. Don't trade this with real capital. The in-sample result is real, but it was found by searching 2,400 configurations on the same year of data. A walk-forward that re-selects parameters using only past data earns +0.7% per trade over 8 months, and that isn't statistically distinguishable from shorting at random. The fixed headline config does better (+7.8%), but it was chosen knowing how these months turned out.

Universe

Every USDT perpetual Bybit listed in the 365 days to 2026-09-28, including ones since delisted. Classification uses Bybit's own instrument type, not a ticker list.

376
USDT perps listed in the window (6 since delisted)
356
With ≥ 14 days of history
114
Crypto-native, the short universe (6 delisted)
242
Stocks, ETFs and commodities (excluded)

Listings by Bybit instrument type

TypeListings
Stock196
Crypto (innovation zone)71
ETF54
Crypto48
Commodity4
Forex3

Two different asset classes

CohortShort d7→14Short d7→30Win rate d7→30
Crypto-native (114)+6.7%+12.1%77%
Stock / ETF / commodity (242)+0.2%−0.5%48%

Tokenised stocks and ETFs don't drift down after listing, so they're excluded. An earlier version classified by ticker list and counted 46 of them (AAL, BRKB, SNDK, TMF, XAU…) as crypto.

The naive short: right direction, untradeable path

The original thesis: market makers support new listings for a while, then withdraw, and price bleeds. Shorting at the day-7 close does make money on average, if you can survive the ride.

Short at day 7 close, all perps

HorizonAverageWin raten
Short day 7 → day 14+2.3%54%356
Short day 7 → day 30+3.6%57%328
Short day 7 → day 90+18.1%80%178

Price only, no costs or funding.

Share of shorts that never hit a stop (day 7→14)

Stop distanceCrypto surviveStock/ETF survive
5%4%44%
10%19%69%
15%30%81%
20%37%92%
30%61%98%

Median adverse move before day 14: crypto 25%, stock/ETF 6%. A time-based entry can't be risk-managed.

Volume does fall after launch week. The median week-2/week-1 turnover ratio is 0.56 for crypto (stock/ETF: 0.56). That's consistent with the market-maker story, though open interest isn't measured.

The condition-based short (in-sample)

Wait for the pump to exhaust instead of shorting on a date.

Entry: all true at the daily close

  1. ATR(14) below its 20-day average: volatility has contracted
  2. Close below the prior day's typical price (H+L+C)/3
  3. Close below EMA50
  4. Still more than +30% above the first full day's open

Execution

  • Short at the next open; symbol at least 50 days old
  • Stop at entry + 3 × ATR(14). It is recomputed daily and only moves down, but it's anchored to the entry, so it never locks in profit.
  • Exit after 30 days; one trade per symbol
+14.8%
Per trade, net
2.33
Profit factor
73%
Win rate
22
Trades
+35.5%
Average win
−40.7%
Average loss

Where the return comes from (per trade)

Price move+17.48%
Funding−2.50%
Fees and slippage−0.21%
Net+14.76%

Funding matters. Shorts on hyped new listings often pay funding. On RIVER, funding of −40% turned a +7.5% price gain into a −33% loss.

Neighbouring configurations

ConfignWin rateAvg winAvg lossPer tradePF
ATR3 Pump30 Hold30 (headline)2273%+35.6%−40.7%+14.8%2.33
ATR3 Pump30 Hold452167%+36.1%−53.6%+6.2%1.34
ATR4 Pump30 Hold302273%+35.6%−52.0%+11.7%1.82
ATR3 Pump35 Hold301974%+38.8%−38.8%+18.4%2.80
ATR3 Pump25 Hold452167%+37.9%−54.5%+7.1%1.39
ATR3 Pump30 Hold30 NoATR2475%+35.2%−53.1%+13.2%1.99

Rank #3 of the grid-search configs with ≥ 20 trades. Average adverse excursion 32%; 18% of trades stopped out; median trade +21.8%. Bootstrap of these trades (10,000 runs): 5th–95th percentile +0.8% to +27.9%, P(mean > 0) = 96%. That bootstrap ignores the parameter search, which is why the walk-forward below exists.

Filter experiments (ATR 4 / pump 20 / hold 60 base)
VariantnPer tradePFStopped
Base50+7.4%1.3932%
+ EMA2050+4.1%1.1934%
+ EMA5050+7.4%1.3932%
+ EMA20+5050+4.1%1.1934%
+ Vol>1.5xSMA2012+3.0%1.1442%
+ Vol>2.0xSMA206too few trades (< 10)
+ Distribution Day9too few trades (< 10)
+ Dist Day Vol>2x4too few trades (< 10)
EMA50 + Dist Day9too few trades (< 10)
EMA20 + Dist Day9too few trades (< 10)
EMA50 + Vol>1.5x12+3.0%1.1442%
No ATR Contraction59+5.6%1.2634%
No ATR + EMA5059+5.6%1.2634%
Pump=15 + EMA5052+2.2%1.1033%
Pump=25 + EMA5049+7.1%1.3633%
Pump=30 + EMA5047+10.9%1.6030%

Volume-spike and distribution-day filters leave too few trades. EMA20 hurts. All figures net of costs and funding.

Walk-forward: the honest test

Each month from 2026-02 to 2026-09, pick the best of 2,400 configurations using only trades that had already closed, then trade that pick, untouched, for the month.

Selection ruleOOS tradesPer tradePFRandom nullp-value
Best past mean33+0.69%1.05−4.10%0.14
Best past mean − 1 SE33+0.69%1.05−4.31%0.14
Headline config, fixed hindsight16+7.79%1.53−7.36%0.02
Out-of-sample return by month, config re-selected monthly
Average net return of the trades entered each month. Hover or focus a column for the config that was picked.
Cumulative out-of-sample P&L
Sum of net per-trade returns in % (one unit per trade), by exit date.
Config re-selected monthlyHeadline config fixed (hindsight)
Month-by-month picks (table view)
MonthTraining tradesConfig pickedOOS tradesPer trade
2026-0217ATR 2× · pump 10% · hold 15d · 3/symbol · no trend filters14−14.6%
2026-0315ATR 4× · pump 25% · hold 15d · 3/symbol · ATR contraction4−20.1%
2026-0415ATR 2× · pump 40% · hold 15d · 2/symbol · EMA503+18.5%
2026-0515ATR 4× · pump 15% · hold 30d · 1/symbol · EMA503+9.9%
2026-0616ATR 4× · pump 15% · hold 30d · 1/symbol · EMA504+19.0%
2026-0718ATR 4× · pump 40% · hold 15d · 1/symbol · EMA503+24.7%
2026-0822ATR 4× · pump 20% · hold 30d · 1/symbol · EMA502+36.4%
2026-0921ATR 4× · pump 40% · hold 40d · 1/symbol · EMA500—

Is it just the minimum-sample choice?

Requiring more past trades before a config can be picked doesn't rescue it. Across these settings, out-of-sample return per trade ranges from −7.0% to +5.8%, and only 5 of 10 variants are positive.

Min. training tradesBest meannBest mean − 1 SEn
10+0.8%30+2.3%32
15+0.7%33+0.7%33
20−3.7%37−4.3%35
25+5.8%23−4.4%34
30−7.0%32−6.4%34

Reading it. The optimiser jumps between very different configs from month to month, which is what fitting noise looks like. The worst months are the earliest, when there were the fewest past trades to learn from; later months are positive, but with this few trades that can't be separated from a favourable stretch. The fixed headline config's +7.8% (p = 0.02) is optimistic by construction: it was picked with knowledge of this whole period.

Headline config trade log

All 22 in-sample trades, net of costs and funding, sorted by result.

SymbolEntryPriceFundingNetMax adverseDaysExit
LAB2026-07-13+71.3%+1.1%+72.2%12%3030-day exit
CLO2026-01-26+81.3%−9.9%+71.2%0%3030-day exit
TRUTH2025-11-29+58.7%+1.6%+60.1%18%3030-day exit
RAVE2026-04-29+48.5%+7.7%+56.0%25%3030-day exit
MAGMA2026-07-17+40.6%+1.3%+41.7%15%3030-day exit
GIGGLE2025-11-30+41.1%+0.1%+41.0%38%3030-day exit
PLAYSOUT2026-06-11+40.2%+0.9%+40.8%5%3030-day exit
SKR2026-03-23+26.9%+0.3%+27.0%3%3030-day exit
US2026-08-23+23.9%+1.0%+24.7%1%3030-day exit
CFG2026-06-02+24.4%+0.2%+24.4%9%3030-day exit
BSB2026-06-02+19.8%+2.5%+22.0%148%3030-day exit
ALLO2026-08-09+22.6%−0.7%+21.6%5%3030-day exit
COAI2025-11-23+16.1%+3.4%+19.3%36%3030-day exit
PRL2026-05-20+24.9%−8.8%+15.8%16%3030-day exit
PIEVERSE2026-01-17+15.0%+0.8%+15.5%35%3030-day exit
KITE2026-03-17+33.1%−17.4%+15.5%22%3030-day exit
NIGHT2026-01-30−7.1%+0.2%−7.1%20%3030-day exit
RIVER2026-02-08+7.5%−40.4%−33.1%90%3030-day exit
CYS2026-02-22−43.3%+0.1%−43.4%43%17stop
UAI2026-04-13−46.4%−0.3%−46.9%46%9stop
TRIA2026-04-19−51.2%+1.3%−50.1%51%14stop
BEAT2026-07-09−63.2%+0.0%−63.4%63%19stop

Other ideas tested

Go long the pump, then flip short on a distribution day (down day, volume ≥ 1.5× average, close in the lower third).

LegnWin ratePer tradePF
Long the pump12336%+9.5%1.69
Short after distribution11544%−8.0%0.71
Combined23840%+1.1%1.05

Gross price returns (this experiment has its own simulator, without costs or funding). The long leg (+9.5%) is more interesting than the short leg (−8.0%). The distribution-day short enters too late. The long leg hasn't been walk-forward tested.

What next

  1. No real capital. The in-sample edge doesn't survive honest parameter selection.
  2. Freeze one config and paper-trade it forward. Only signals after the freeze count as evidence. At about 1.8 signals a month, 30–50 of them takes one and a half to two years.
  3. Add a funding filter or exit. Funding can erase a winning short.
  4. Use a longer history. A one-year window is the main limit on statistical power.
  5. If ever live: stops sit 40–80% away, so size by stop distance and keep leverage around 1× to avoid liquidation before the stop.