H-033: Gate.io GRAM 'smart wallets': deposits of $500+ by wallets picked on 2023-24 foretell 7-day drops better than random wallets
Fail Family: flows · Status: rejected
Do 'smart money' GRAM deposits on Gate.io predict drops? The idea did not beat simply buying the coin and holding it. It is rejected.
Result
36 smart wallets of 183; from 2025-01-01 their 144 deposits were followed by a mean 7-day log return of -0.0050 vs -0.0196 for random matched wallets (p = 0.768, need p < 0.05); exit rule Sharpe -1.80 vs hold -0.71
| Selection 2023-03-13 – 2024-12-24 | 183 eligible wallets, 36 smart, 1565 events |
|---|---|
| Mean 7-day move after deposits (selection): smart / all | -3.85% / +1.34% |
| Test from 2025-01-01 | 144 smart-wallet events from 12 wallets |
| Mean 7-day move after deposits (test): smart / random wallets | -0.50% / -1.94% |
| p (random wallets, matched) | 0.768 |
| Exit rule vs holding: Sharpe | -1.80 / -0.71 |
| Exit rule vs holding: CAGR | -38.2% / -56.1% |
| Exit rule vs holding: max drawdown | -59.1% / -79.3% |
| Exit rule: time in the coin | 16% |
Why it might work
Same question as H-032 on Gram (GRAM, prev. Toncoin), asked for by F1DevBin after H-032 failed. On TON, Gate.io takes deposits on one shared wallet (0:b496b8f7...60bc281f, labelled 'Gate.io' and whitelisted by tonapi, active since 2022-11-04): users send GRAM straight to it with their Gate account number as a text comment, so there are no sweeps and a 'wallet' is the sending address (crl data ton: public toncenter v3 index, text-comment transfers that did not bounce, addresses stored only as short SHA-256 ids; the deposit wallet itself and other Gate.io-labelled accounts are dropped, as are the owner's addresses if he names any). Differences from H-032, all decided before any outcome: windows follow GRAM's price history on Gate.io (from 2023-03-13): selection 2023-03-13 to 2024-12-24, test from 2025-01-01; no sweep rule. Size: F1DevBin found 10 coins too small to matter, so an event is a transfer worth at least $500 at the close of its day (when the event is decided). The size was picked by a rule written down before the data were counted (scripts/h033_threshold.py): the highest of $10 000, $5 000, $1 000, $500 with at least 75 wallets having 5+ selection events and at least 300 test-period events for them; counts only, no return was computed ($10 000: 81 wallets and 109 test-period events; $5 000: 88 and 135; $1 000: 137 and 219; $500: 183 and 342, so $500). Services: wallets with more than 100 transfers of 1+ GRAM in a calendar month of the selection period. Everything else as in H-032: events of one wallet within 7 days join; outcome = log return of GRAM (Gate.io daily closes) from the close of the event's day to 7 days later; 'smart' = lowest 20 % mean outcome among wallets with 5+ selection events (need 25+); test: their events' mean outcome (need 30+) vs 1000 random sets matched on selection-event counts, pass if p < 0.05 (one-sided); exit rule with Gate's 0.2 % fee + 0.2 % slippage vs holding. GRAM only: wallet-level data is per chain (exception to the all-coins rule). Expectation written before any result: like H-032, the 'smart' wallets were lucky, not informed.
Verdict history
- 2026-10-28: Fail (this engine, data to 2026-10-27)
Pre-registered specification
kind: wallet_deposits
asset: GRAMUSDT
universe:
- GRAMUSDT
deposit_wallet: 0:b496b8f73b469ff06a2d89e62c4d70a103fc228c821537b49dff092060bc281f
dataset: gram_gate_deposits.csv.gz
filters:
min_value: 1
min_usd: 500
swept_before: false
merge_days: 7
max_monthly: 100
selection:
start: '2023-03-13'
end: '2024-12-24'
min_events: 5
horizon_days: 7
smart_share: 0.2
min_wallets: 25
test:
start: '2025-01-01'
min_events: 30
'null':
method: random_wallets_matched
n: 1000
buckets:
- 5
- 6
- 8
- 11
- 16
- 26
significance:
alpha: 0.05
rule:
exit_days: 7
fee: 0.002
slippage: 0.002