Knowledge · Research · Cross-Sectional Momentum for Botty - market-neutral: buy the winners, short the losers (measured, with visualizations)

Cross-Sectional Momentum for Botty - market-neutral: buy the winners, short the losers (measured, with visualizations)

Strategy analysis 2026-06-20 10 sources
The third structural edge of the dead-end initiative (after funding-carry live + VRP harvest infra-gated): every week go long the relative winners (top tercile) and short the losers (bottom tercile), dollar-neutral across 20 HL coins. Market-neutral (BTC beta ~0) - we do NOT forecast direction, we bet on relative strength (the momentum anomaly). Tamed with inverse-vol + vol-targeting + crash filter (sit out bear-market rebounds). It PASSED the cross-universe gate that killed donchian (generalizes across disjoint coin sets), PBO 0.27 solid, walk-forward time-stable; Deflated Sharpe 0.71->0.93 with crash filter (just under 0.95 = real but borderline). Backtest ~+42%/yr filtered, honest expectation ~15-25%/yr market-neutral. Runs as a monitor in the bot, armed after wallet top-up. With principle visualizations.
  • THE IDEA: market-neutral cross-sectional momentum. Every week rank 20 coins by 30d return, long the top tercile / short the bottom tercile, equal size on both sides (dollar-neutral). The bet = relative strength is more persistent than absolute market direction (the momentum anomaly since 1993). BTC beta ~0 measured = truly directionless. [[cross_sectional_momentum]].
  • WHY DIFFERENT from our dead ends: no BTC directional forecast (that died at every gate), but relative + market-neutral. Runs on HL perps - no capital floor like with options ([[optionen_fuer_botty]]), can be run arbitrarily small.
  • THE TAMING (3 layers): (a) inverse-vol weighting per leg (risk parity, wild coins smaller); (b) vol-targeting on the book (constant risk); (c) crash filter - the most important: momentum crashes in bear-market rebounds (the shorted losers explode), the filter goes flat when the market basket is >20% below its 90d high. Effect: 2026 crash -30%->+3%, maxDD -21%->-11%, Sharpe ~1.0->~2.2.
  • CROSS-UNIVERSE GATE PASSED (the one that killed donchian): frozen config on disjoint universes - LARGE-20 Sharpe +0.96, MIDCAP-20 (disjoint) +0.43, SPLIT-A +0.73, SPLIT-B +0.48, ALL positive + market-neutral. Generalizes = a real effect, not a coin fit. The second structural edge after funding to pass a generalization gate.
  • PBO 0.27 (solid <0.3) + walk-forward time-stable (mature folds IS=OOS=0.18, no overfit degradation, selection stable at lb30). The strategy family is real, not selection overfit.
  • DEFLATED SHARPE borderline: 0.71 (raw) -> 0.93 with crash filter, JUST under 0.95. The absolute Sharpe is too small to be safely significant after the 40-trials correction. Honestly: real but borderline - a diversifying sleeve, not a home run. The 0.93 is itself slightly optimistic (the filter search is not captured in N=40).
  • EXPECTATION: backtest LARGE-20 filtered ~+42%/yr (Sharpe 2.2, maxDD -11%), honest forward expectation ~+15-25%/yr market-neutral (haircuts: midcaps half as strong, slippage, DSR borderline). The value = UNCORRELATED to funding-carry + directional = genuine diversification.
  • LIVE STATUS: execution/xsectional/ built (signal/broker/strategy_runner), config XSECT_MOMENTUM on free wallet 2, its own UI tile+dossier. Runs as a MONITOR (shows target book + crash regime, 0 orders). Armed after wallet top-up (~$300-500) + env XSECT_ARMED. Evidence: backtesting/xsectional_{v2,universe_gate,wf_pbo,crash_filter}.py.
Run it as a diversifying market-neutral sleeve, NOT as a home run: honest expectation ~15-25%/yr, DSR borderline (~0.93). The value is the lack of correlation to funding-carry + donchian, not the absolute return.
Run ONLY tamed + crash-filtered (inv-vol + vol-target + crash filter), NEVER raw: the crash filter (flat when the market is >20% below its 90d high) is the difference between Sharpe ~1 with -30% years and Sharpe ~2.2 with a capped tail.
Start small on HL perps (no capital floor): wallet 2 at ~$300-500, 2x gross leverage lifts the 20 coins above HL minimum sizes. Market-neutral = low directional risk, but liquidation on an extreme single-coin move is not 0. Supervise the first 1-2 real rebalances.
Crash filter ONLY with a rolling window (90d high), NEVER all-time high: from the all-time high the crypto basket sticks flat for years = a bull-timing trap (82-88% flat, fair-weather Sharpe). The rolling variant recovers after a real drawdown (trap caught, 2026-06-20).
Respect the DSR threshold honestly: 0.93 < 0.95 does not mean 'dead' but 'not proven'. Live paper performance reads out slowly (a weekly bet) - the monitor validates EXECUTION quickly (1-2 rebalances), the performance remains the 278-week backtest.

The idea in one sentence

Every week buy the relative winners and short the relative losers - equal amounts on both sides, so that the market's direction is irrelevant to us. We are not betting "will crypto rise?", only "do the strong coins keep outperforming the weak ones?".


1. How it works concretely

Imagine you watch 20 cryptocurrencies at once (BTC, ETH, SOL, ...). Every week you do the same thing:

  Step 1 - SORT (by 30-day return)
  strongest  ┌──────────────────────────────────────────┐  weakest
             │ SOL  ETH  LINK │ BTC XRP ADA … │ DOGE LTC FIL │
             └──────┬─────────┴───────┬───────┴──────┬───────┘
                    │                 │              │
  Step 2 - BET
             ┌──────▼──────┐   (middle:│      ┌──────▼──────┐
             │  BUY long    │   ignore) │      │ SELL short    │
             │ top tercile  │                  │ bottom tercile│
             │ ~7 coins     │                  │ ~7 coins      │
             └─────────────┘                   └──────────────┘
                    └──────── equal $ amount ───────┘
                         = dollar-neutral

The bet behind it is called momentum: "what has done well recently usually keeps doing a bit better - and what has done badly usually stays weak." This is the best-documented anomaly in financial science (since 1993, robust across 58 markets).


2. The trick: the market's direction is irrelevant to us

Because we are equally long and short, a crash cancels out:

  Market CRASHES −30%                  Market PUMPS +30%
  ───────────────                      ───────────────
  Long leg:  −30%  (loss)              Long leg:  +30%  (gain)
  Short leg: +30%  (gain!)             Short leg: −30%  (loss)
  ─────────────────────────            ─────────────────────────
  net:       ~0  ✓                     net:       ~0  ✓

  What remains: ONLY the difference winners−losers.

Measured: BTC beta ~0 - the overall market is truly irrelevant to us. This is the fundamental difference from everything we tried before on BTC (and that died at the gates): we do not forecast direction.


3. The taming - three layers against the one weak point

Raw, the strategy is real but fragile. Three fixes make it tradeable:

(a) Inverse-vol weighting (risk parity). Wild coins get less weight, so that no single coin drives the risk.

(b) Vol-targeting. The whole book is scaled inversely to its own volatility -> constant risk.

(c) The crash filter - the most important. Momentum has one deadly weak point: the bear-market rebound. The market falls deep, then snaps back sharply - and the shorted losers explode upward:

  MOMENTUM CRASH (without filter)
       Price
        │       ╱╲  ← sharp rebound: the losers
        │      ╱     (which we SHORT) shoot up
        │  ╲  ╱            → our short leg bleeds
        │   ╲╱
        └──────────────► Time
            ↑ bear low

  THE FILTER: if the market basket is >20% below its 90-day high,
  the WHOLE BOOK goes flat - we simply sit out the rebound.

Effect in the backtest: the 2026 crash year went from −30% to +3%, the maxDD halved (−21% -> −11%), and the Sharpe doubled (~1.0 -> ~2.2).


4. The gauntlet - which gates it passed

We put it through the same tests that killed almost everything else:

  GATE                         RESULT                         Verdict
  ─────────────────────────────────────────────────────────────────
  Anomaly real?                Sharpe +, BTC beta ~0           ✓ yes
  Taming works?                maxDD halved, Sharpe holds       ✓ yes
  Cross-universe gate          MIDCAP +0.43 · SPLIT-A/B + ·    ✓ PASSED
   (killed donchian!)           all disjoint positive            (no coin fit)
  PBO (selection overfit?)     0.27 (<0.3)                      ✓ solid
  Walk-forward (time-stable?)  mature folds IS=OOS=0.18         ✓ stable
  Deflated Sharpe              0.71 → 0.93 with crash filter   ~ JUST <0.95
  ─────────────────────────────────────────────────────────────────

The decisive moment: the cross-universe gate. It tests whether the edge generalizes to other coins or is only fitted to the 20. This is exactly where donchian@2h died (BTC +66% / ETH −11% / SOL −60%). Cross-sectional survives it - the second structural edge after funding-carry to pass a generalization gate.

The only fly in the ointment: the Deflated Sharpe (corrected for the fact that we tried 40 variants) lands at ~0.93 - just under the 0.95 significance threshold. Honestly: real, but borderline.


5. What we honestly expect

Value
Backtest (LARGE-20, filtered) ~+42%/yr, Sharpe ~2.2, maxDD ~−11%
Honest forward expectation ~+15-25%/yr, market-neutral
Why lower midcaps half as strong · live slippage · DSR borderline

The +42% is the favorable backtest reading on the strongest universe. Forward, after all the honest haircuts, we expect ~15-25%/year - and the real value is not the number but that this return is uncorrelated to funding-carry and to the directional strategies. Genuine diversification.


6. Why this strategy is different from our dead ends

  Directional BTC strategies      Cross-sectional momentum
  ───────────────────────────     ────────────────────────
  "will BTC rise?" (forecast)     "winners > losers?" (relative)
  single-asset                    20 coins against each other
  fully exposed to the market     market-neutral (beta ~0)
  died at the cross-asset gate    PASSED the cross-universe gate
  Deribit/capital floor (options) HL perps, run arbitrarily small

It is the embodiment of the whole "out of the BTC-timing dead end" initiative: structural instead of directional, relative instead of absolute. Together with the funding-carry (alive) and the VRP harvest (infra-gated), it is the third proof that the axis is right.


7. Live status

The strategy runs as a monitor in the bot (its own UI tile): it computes the target book live and shows the crash regime, but sends no real orders yet. It is armed after a wallet top-up (~$300-500 on the free wallet 2) + the XSECT_ARMED latch - then it rebalances for real weekly on Hyperliquid perps.

A nice live proof during the build: the crash filter kicked in immediately for real - the market was >20% below its 90-day high, so the strategy held the book flat. Exactly as designed.