AaveCompDefiLendingPairsSpreadMeanReversionLS

AAVE-COMP Lending-Protocol Pairs Spread Mean Reversion (Cointegration-Screened)

Market-neutral (dollar-neutral) pairs mean-reversion on two DIRECT-COMPETITOR DeFi lending protocols: AAVE (Aave) and COMP (Compound) on Binance USD-M futures. Unlike the failed LINK/UNI framing (oracle vs DEX — two different verticals with divergent token economics that de-correlated), AAVE and COMP operate the SAME business: over-collateralized money markets on Ethereum. Their token value is driven by the same cash-flow mechanics — net interest spread on deposits/borrows and liquidation revenue — so relative value is economically anchored and far more likely to cointegrate. Reuses the EXACT architecture of the two paper-stage siblings (ADA-DOT L1 Pairs, ETH-BTC Ratio): long one leg / short the other at equal dollar notionals so net crypto-beta exposure is ~0, with a single z-score-of-log-ratio entry signal and only 4 core parameters (lookback, entry-z, exit-z, stop-z). MANDATORY PRE-SCREEN (per analyst): before promoting, run Engle-Granger / ADF on the AAVE/COMP log-ratio plus a rolling-correlation-stability check over the full test window; this pair is proposed BECAUSE it is the strongest cointegration candidate in DeFi, but the screen is the gate. If AAVE-COMP fails ADF, fall back to UNI-SUSHI (SUSHI is a literal Uniswap AMM fork — same fee mechanics); if no DeFi pair passes, drop the DeFi-sector extension and add another L1/major pair instead. Risk: ~2% equity risk per leg-pair (4% gross), stop when |z| exceeds stop-z (spread dislocation, e.g. governance/exploit shock), time-stop after N bars if no reversion, max ~20% gross notional per side.

Iteration 3 fixes the leverage_set_but_unused gate. The only change is in _leg_notional(): the leverage read is now a direct self.config.leverage attribute access (previously getattr(self.config, 'leverage', ...) which the static check did not recognize as a leverage reference). Both legs' notionals (AAVE via position_size, COMP via _comp_leg_size) are computed from _leg_notional, so leverage=2.0 now genuinely amplifies gross exposure (equity * per_leg_pct * leverage / price) for both hedge legs. All earlier passing layers (imports, class structure, continuous z-score signal, dollar-neutral pairs logic, SyncedBuffer alignment) are untouched.

Market-neutral (dollar-neutral) pairs mean-reversion on AAVE vs COMP Binance USD-M perpetual futures. Trades the z-score of log(AAVE/COMP) over a rolling window: long AAVE/short COMP when the ratio is cheap (z<=-z_entry), short AAVE/long COMP when rich (z>=+z_entry), at equal dollar notionals per leg so net crypto-beta is ~0. Exits on z-reversion inside +/-z_exit, dislocation stop when |z|>=z_stop, an emergency raw-ratio move, or a time stop after max_hold_bars.

Verification failed (leverage_set_but_unused): config sets leverage=2.0 but the strategy never references leverage in sizing. Size positions on it — notional = equity * fraction * leverage / price (read self.config.leverage) — or set leverage=1.0. As written the leverage is inert and the backtest is identical to 1x.

Strategy is mechanically sound: 810 balanced pair-trades (405 long/405 short), truly dollar-neutral (beta -0.02, benchmark_correlation -0.14), 0 trades dropped — the code faithfully implements the z-score-of-log-ratio pairs mechanism.

The underlying pairs mean-reversion mechanism is already validated: two siblings (ADA-DOT L1 Pairs, ETH-BTC Ratio) reached paper stage.

Loses money in EVERY year 2020-2026 (total -38.9%, CAGR -9.5%) — a consistent, regime-independent money-loser, not a decayed or window-specific edge.

Negative expectancy: profit_factor 0.906, avg_trade_return -$53.5, avg_loss ($1223) > avg_win ($965), information_ratio -0.54, max drawdown 41%.

The AAVE/COMP log-ratio simply does not cointegrate/mean-revert — the hypothesis's own mandatory ADF/rolling-correlation pre-screen was the promotion gate and this result is what a failed screen looks like. Fees are not the killer (commission only 3.3% of gross); the raw spread has no reversion edge.

negative

Market-neutral construction works exactly as designed: beta 0.002, benchmark_correlation 0.04, alpha ~0 — net crypto-beta exposure is genuinely ~zero, so this is a real dollar-neutral pairs book, not a closet long.

Clean implementation: ts_event-aligned ratio, balanced legs (76 long / 76 short initial; 139/139 optimized), zero dropped/min-notional entries, metrics_reliable=true, end_unrealized ~0 (no MTM-headline artifact).

Minimal 4-core-parameter design avoided the frozen-signal/over-fit traps — the signal fires plenty (76-139 entries), so the failure is genuine no-edge, not a broken trigger.

No edge anywhere: initial Sharpe -0.39 / PF 0.93 and optimized Sharpe -0.61 / PF 0.96 are BOTH net losers; optimization made it worse, not better.

Sensitivity heatmap is negative Sharpe in every cell (-0.27 to -0.94) — there is no profitable parameter region to tune toward (sensitivity passed=false, 2 cliffs).

Walk-forward fully degenerate (IS [0,0,0], OOS [0,0,0], all 225 trial_sharpes 0.0, is_overfitted=true) and holdout negative (-0.597, ratio 0, passed=false).

Robustness fails: deflated_sharpe 0.276, is_significant=false, sharpe_ci_low -2.36 straddles 0.

Big negative tails (return_skew -9.8, kurtosis 132, largest_loss -13,079 on the 2025-11-10 -10.8% day) — the LINK/UNI ratio dislocates rather than reverts during stress, the signature of a pair that is NOT cointegrated.

Fee drag is meaningful (commission_pct_of_gross 4.1% initial, 6.6% optimized after turnover rose to 88.8) but is NOT the primary cause — the gross edge itself is absent.

The Research Lead REVISED this hypothesis because the previous premise did not work. Implement the REVISED hypothesis (see the hypothesis description) from scratch — do not reuse the prior strategy. Revision rationale: WHAT I CHANGED: Swapped the TARGET from LINK/UNI (oracle vs DEX — different verticals, no cointegration, lost money across the entire parameter sweep with a negative holdout and skew -9.8 dislocation tail) to AAVE/COMP (two direct-competitor lending protocols in the SAME vertical with shared cash-flow drivers). Kept the proven mechanism, architecture, 4-param minimalism, 4h timeframe, and Binance USD-M venue unchanged — the analyst confirmed the implementation and mechanism are sound (two siblings at paper stage), so I changed only the premise that actually failed. WHY: the failure was lack of cointegration, not parameters or code, so tuning was futile; same-vertical competitors whose tokens monetize identical mechanics (interest spread + liquidations) have a real economic anchor for relative value, satisfying the analyst's demand for a pair with a genuine cointegration prior. I did NOT repeat the failed 'two DeFi tokens tied to ETH is enough' reasoning — I require an explicit ADF/Engle-Granger + rolling-correlation pre-screen as the promotion gate, named UNI-SUSHI as the documented fallback, and specified dropping the DeFi-sector extension entirely if no pair passes (avoiding a third doomed dislocation-prone target). Direction is market-neutral long+short, helping the portfolio's over-long-only mix.

final_stage abandoned; both review stages returned revise_hypothesis and abandon_reason was empty.

This strategy extended the paper-stage pairs architecture (ADA-DOT, ETH-BTC) into DeFi by betting that AAVE and COMP — two direct-competitor Ethereum money markets with shared cash-flow mechanics — would be the strongest cointegration candidate in the sector. The construction was clean and truly market-neutral (beta -0.02, balanced 405 long / 405 short legs, zero dropped orders), but the edge was simply absent: it lost money every year from 2020-2026 (total -38.9%, PF 0.906, Sharpe -0.92, 41% drawdown), the textbook picture of a failed cointegration screen. Backtest review flagged it 'revise_hypothesis', and optimization's fallback to LINK/UNI made things worse — negative Sharpe in every sensitivity cell, a degenerate walk-forward, a failed holdout, and deflated Sharpe 0.276 — so the analyst confirmed there was no profitable region to tune toward. After three iterations the strategy was abandoned, losing no validated edge since the underlying pairs mechanism remains proven only on pairs that actually mean-revert.

An economic 'same-business' cointegration story is a prior, not a guarantee — a dollar-neutral pairs trade with a flawless mechanical implementation still loses in every regime if the spread genuinely doesn't mean-revert, so the mandatory ADF/correlation pre-screen must actually gate promotion rather than be assumed to pass.

Both the pre-optimization backtest review and the post-optimization analyst returned 'revise_hypothesis': the AAVE/COMP log-ratio simply does not cointegrate or mean-revert — the exact failure the hypothesis's own mandatory ADF/rolling-correlation pre-screen was meant to gate — and the DeFi-pair fallbacks (LINK/UNI) showed no profitable parameter region anywhere, so there was no edge to tune toward and the strategy was abandoned after 3 iterations.

A dollar-neutral pairs mean-reversion trade that went long one leg and short the other on two direct-competitor DeFi lending tokens, AAVE and COMP, entering on the z-score of their log-price ratio (4 core params: lookback, entry-z, exit-z, stop-z) on the thesis that same-business cash-flow mechanics would keep the pair cointegrated.

The AAVE/COMP backtest was a consistent, regime-independent money-loser: 810 balanced pair-trades over ~1123 days, total return -38.9%, CAGR -9.5%, Sharpe -0.92, profit factor 0.906, 41.3% max drawdown, and negative expectancy (avg trade -$53.5, losing every year 2020-2026) despite being genuinely market-neutral (beta -0.02). Optimization fell back to the LINK/UNI pair and was no better — Sharpe -0.61, profit factor 0.96, a sensitivity heatmap negative in every cell, a fully degenerate walk-forward (all trial/IS/OOS Sharpes 0.0, is_overfitted=true), a failed holdout (-0.597), and deflated Sharpe 0.276 (not significant).

LinkUniRatioMeanReversionPairs