What this is: A BTC + ETH + cash timing strategy driven by a 21-model conviction ensemble. Every week the strategy decides: are we in BTC, are we in ETH, are we leveraged, or are we in cash. It never trades alts. Maximum 2 simultaneous positions. ~41% of historical days are full cash.
Why no alts? Five alt-trading sleeves were tested empirically and dropped: ALT_XSMOM (OOS SR -0.21), ALT_DIVERS (0.16), BROAD_XSMOM top-5-of-100 (-0.26, -60% MDD), QUINTILE_SPREAD (-0.58, -67% MDD), DISPERSION_TIMED (0.31). Daily-bar alt momentum systematically picks parabolic alts right before they reverse — the dead coins in our universe (LUNA, FTT, etc.) were exactly the kind of "top momentum" names this would have bought.
Objective: ~50% CAGR with -32% MDD on the BTC+ETH+cash space, executable manually on OKX (leveraged tokens) or Hyperliquid (perpetuals). Long-only. Weekly rebalance Wednesdays.
How the timing decision is made: 21 weakly-correlated sleeves each output a per-coin target weight (mostly 0 or some BTC weight, occasionally ETH). The aggregated weights are sleeve-vol-targeted, risk-parity-blended, consensus-gated, and weekly-snapped. A daily vol-target overlay scales total exposure. A conviction-scaled leverage curve (1×→3.2×) amplifies high-agreement weeks. A vol-of-vol regime gate cuts leverage when vol is expanding. A self-stop halves exposure on drawdowns over -4% in 30 days.
21 timing models (every model outputs only BTC and/or ETH weights):
The sleeves are deliberately heavy on BTC because empirically, 11 years of crypto data show BTC has the cleanest signal-to-noise ratio for daily-bar trend, breakout, and regime models. ETH adds a small uncorrelated layer; alts add noise.
Portfolio overlays (1-day lagged):
scale = ((conv − 0.15)/0.55)1.1 × 1.35; below 0.15 sleeve agreement, full cash.Why a 50-coin signal universe if we only trade BTC/ETH?
35 survivor coins (signal-only): BTC, ETH, SOL, ADA, DOGE, LTC, BCH, XRP, LINK, DOT, AVAX, ATOM, XLM, TRX, ALGO, BNB, XMR, DASH, EOS, NEO, ETC, ZEC, HBAR, VET, FIL, THETA, XTZ, QTUM, MKR, BAT, IOTA, DCR, ZRX, LRC, ICX.
15 dead coins (signal-only): LUNA1 (Terra collapse), USTC (depeg), FTT (FTX collapse), MATIC (migrated to POL), UNI (delisted), XEM (NEM -99%), BTG (51%-attacked), STEEM, SC, OMG, WAVES (USDN collapse), CELR, ANT, GRT, KNC.
Of these 50 coins, only BTC and ETH ever appear in the actual position book.
After observing that the strategy only ever holds BTC and/or ETH, we ran two systematic tests of "scan a wide universe and allocate to alts when conditions are right." Both failed empirically. Test results below — committed to the feature branch for full audit.
Per-coin scoring on 5 independent signals: long-trend (price>200MA AND 50MA>200MA), short-mom (21d>5%), persistence (63d AND 252d positive), not-overbought (price < 0.93×21d-high — filters parabolic tops), stable-vol (21d vol < 1.5× 252d median). Strict eligibility: 252d minimum history, not in >25% DD, positive 365d return, BTC>200MA AND BTC 63d-mom > 5%. Top-K by 63d momentum among coins with confluence ≥ min_signals.
| config (top_k, min_signals) | Standalone OOS SR | Standalone Full SR | Ensemble OOS SR | Ensemble CAGR |
|---|---|---|---|---|
| k=3, min=3 | +0.17 (50-univ) / -0.35 (111-univ) | 0.66 / 0.50 | 1.12 | 52.5% |
| k=5, min=4 | -0.11 / -0.09 | 0.64 / 0.53 | 1.09 | 50.2% |
| k=7, min=5 | -0.98 / -0.69 | -0.04 / 0.03 | +1.16 (best) | 42.7% |
| k=10, min=3 | +0.09 / +0.04 | 0.94 / 0.89 | 1.10 | 52.0% |
| BASELINE (no scan) | — | — | 1.13 | 51.4% |
Verdict: All 12 configurations have negative or near-zero standalone OOS SR. The best ensemble config (k=7, min=5) gives +0.03 OOS SR but loses 8.7% CAGR (NAV halves from 124× to 62×). Net-negative trade-off.
Eight new sleeves: SOL_VM, BNB_VM, AVAX_VM, LINK_VM, ADA_VM, XRP_VM, LTC_VM, DOGE_VM. Each is a dedicated vol-managed trend follower on one coin (same logic as BTC_VM and ETH_VM). All BTC-regime gated (only fire when BTC>200MA AND BTC 63d mom > 0).
| Coin sleeve | Standalone Full SR | Standalone OOS SR |
|---|---|---|
| SOL_VM | 0.86 | +0.36 |
| BNB_VM | 0.78 | +0.78 |
| AVAX_VM | 0.51 | +0.12 |
| LINK_VM | 0.55 | +0.08 |
| ADA_VM | 0.60 | +0.49 |
| XRP_VM | 0.39 | +0.54 |
| LTC_VM | 0.35 | -0.22 |
| DOGE_VM | 0.25 | -0.03 |
Ensemble impact (with all 8 alt sleeves added):
| Full SR | OOS SR | CAGR | MDD | NAV | 2025+ | |
|---|---|---|---|---|---|---|
| Baseline (BTC+ETH only) | 1.47 | 1.13 | 51.4% | -32.1% | 124× | -3% |
| + 8 alt sleeves | 1.17 | 0.92 | 30.7% | -35.8% | 22× | -14% |
| Δ | -0.30 | -0.21 | -20.7% | -3.7% | -102× | -11% |
Verdict: Adding 8 alt sleeves to the ensemble DROPS OOS SR by 0.21 and 80% of NAV. The 2025+ year goes from -3% to -14% because alts crashed harder than BTC in 2025 chop. Cherry-picking just the OOS-strong sleeve (BNB_VM at 0.78) would be look-ahead bias.
sleeves.py for audit.
data/crypto/ (yfinance via scripts/expand_universe.py).python live_signal.py --venue okx --capital <USD> after market close.hyperliquid-python-sdk).