EthVwapReversionLS

ETH Intraday VWAP-Reversion, Long-Short — Gap-Safe Port of the Deployed SOL Survivor (Single-Instrument ETHUSDT.BINANCE Perp, 15m Bars, Fade MODERATE Deviations from Rolling VWAP, Trend-Gate OFF, Frequently-Firing, Capped Notional + 1x Leverage, 3-Parameter)

A LONG-SHORT intraday mean-reversion that deliberately PORTS the exact recipe of the factory's ONE surviving price strategy — the deployed SOL USD-M Intraday VWAP-Reversion (15m, fade moderate deviations, trend-gate OFF, frequently-firing, low-parameter, at paper_stage) — to ETH, the instrument that keeps its fee-clearing volatility WITHOUT the catastrophic-gap history that liquidated the XRP reversion sibling. This is explicitly OUTSIDE every family the meta-learner just declared dead: it is NOT a majors momentum-rotation (L32 exhausted — my three pending rotations are now duplicate-in-spirit), NOT a funding-carry / negative-funding basket (L33 — carry never clears its execution cost), NOT a quarterly/COIN-M basis (L34 data-unavailable), NOT options (fragile/blocked). Intraday reversion is the single price mechanism with a live survivor and is not on any exhausted list. It fills the under-represented SHORT horizon (8.7%) and LONG-SHORT direction (14.3%). The XRP-blowup death mode is engineered out by three specifics: (1) ETH has never gapped anywhere near -100% intraday (unlike XRP's Nov-2020 pump/dump), (2) FIXED 1x leverage, (3) a HARD notional cap so even a 30% adverse gap costs <5% of equity — no liquidation is possible. Only 3 parameters, matching the survivor's simplicity against the overfit graveyard.

Iteration 2 targets the exact Layer-4 QA diagnosis: a well-sampled run (1416 trades) with negative post-fee edge (avg_trade_return_pct -0.034%, PF 0.70) driven by an INVERTED reward/risk — win_rate 0.65 but avg_loss ~2.6x avg_win, because the only loss controls were a 3% price stop (~8 sigma at 15m) and an 8h time-stop, so the tight exit_z=0.25 banked ~1-sigma wins while losers ran. The fix is structural and minimal: (1) a z-based adverse stop (stop_z=2.8) becomes the binding loss control, cutting a failed fade at ~(stop_z-entry_z)=1.0 sigma instead of ~8 sigma — this flips the loss/win ratio from ~2.6 to ~0.65 (win ~1.55 sigma vs loss ~1.0 sigma), which yields positive gross expectancy even at a 50% win rate (+0.275 sigma) and clearly positive at the observed 0.6-0.65; (2) entry_z raised 1.2→1.8 so each captured reversion (~1.55 sigma of band width) is larger relative to the ~0.10% round-trip taker fee, and deeper deviations revert more reliably; (3) the time-stop lengthened to 12h so genuine reversions can complete while the z-stop caps the downside. The 3% price stop is retained only as a catastrophic gap backstop, preserving the un-liquidatable 15%-notional 1x design (a 30% gap still costs <5% of equity). All other code is untouched, so earlier layers stay green: imports/structure unchanged (Layer 1), calculate_signal is byte-identical so the signal still varies (Layer 2), and entry_z=1.8 still fires thousands of times over the 15m history so the sandbox produces ample trades (Layer 3, no verification_loop). The change adds one risk parameter (stop_z) — justified because the reported failure was precisely uncontrolled loss magnitude — while keeping the model lean against overfitting. This is an honest attempt to convert the mechanism to positive expectancy; if the deeper-fade + hard-z-stop still cannot clear fees on ETH 15m, the result will show it cleanly, but the reward/risk asymmetry that caused this specific failure is now removed.

Iteration 2's edge is confirmed negative on a decisive 6,130-trade sample: avg_trade_return_pct -0.028% (below the 0.15% fee floor AND below zero), profit_factor 0.774, total_return -94.5%, Sharpe -0.87, max_drawdown 94.5%. The iter-2 fix explicitly added a z-based adverse stop to repair the inverted reward/risk (was win 0.65, avg_win ~$79 vs avg_loss ~$207) — it FAILED: avg_win $86.9 vs avg_loss $169.1 remains ~1.9x inverted and per-trade expectancy stayed negative. Every year 2020-2026 and nearly every month is negative, so there is no profitable regime — this is an absence of directional edge, not a tunable window. The hypothesis's premise that the deployed SOL VWAP-reversion recipe ports to ETH is falsified; this is the same VWAP-fade mean-reversion class that has died repeatedly this session (SOL/ETH/XRP/DOGE/BNB/LINK) with the identical negative-expectancy, inverted-RR signature. Optimizing entry_z/exit_z/stop_z cannot flip negative per-trade expectancy — the developer already spent iteration 2 tuning exactly those loss controls without success. Abandon rather than spend 2 hours optimizing.

Long-short intraday VWAP mean-reversion on ETHUSDT.BINANCE USD-M perp, 15m bars, pure OHLCV, 1x leverage. It computes a rolling volume-weighted VWAP over vwap_window (48 = 12h) bars and the volume-weighted band std, and measures the close's deviation as a z-score. It fades symmetrically (trend-gate OFF): short when z > entry_z (1.8), long when z < -entry_z. Exits, in priority order: (1) a Z-BASED ADVERSE STOP — cut when the deviation widens against the position past stop_z (2.8) sigma, capping each loss at ~(stop_z - entry_z) = ~1 sigma; (2) a 3% catastrophic price backstop for gap safety; (3) a side-aware reversion take-profit once the z-score collapses to/through the mean band (LONG exits at z >= -exit_z, SHORT at z <= exit_z); (4) a 12h time stop. Iteration 2 added the z-based stop and raised entry_z specifically to fix the QA-diagnosed inverted reward/risk (small wins, big losses). position_size applies a HARD notional cap of 15% of equity at 1x, so a 30% adverse gap costs <5% of equity.

Verification failed (Layer 4 — QA review): - Decisive negative post-fee edge on a well-sampled run. The Layer-3 sandbox produced 1416 trades (metrics_reliable=true) with avg_trade_return_pct = -0.034% — below the ~0.10% round-trip taker cost and the 0.15% futures viability floor. Supporting: profit_factor 0.70, total_return -31.3%, Sharpe -2.43 with 95% CI [-4.08, -0.82] entirely below zero, max_drawdown 31.4%. The failure is structural, not statistical: win_rate 0.65 but avg_win $78.7 vs avg_loss $207.4 (~2.6x) — the tight exit_z=0.25 reversion take-profit banks small moves while the 3% hard stop / 8h time-stop lets losers run, so the fade wins often but loses big and nets negative. This is the exact inverted-reward/risk signature of the single-name pure-OHLCV VWAP/mean-reversion class that has repeatedly died this session. The 'gap-safe port of the SOL survivor' premise did not transfer to ETH. - The code correctly and faithfully implements the stated hypothesis (ETHUSDT.BINANCE 15m perp, long-short symmetric VWAP z-score fade, trend-gate off, 3 tuned params, 1x leverage, hard 15% notional cap, hard stop + time stop). Signal polarity, API usage, the once-per-bar counter, and zero-volume/zero-denominator guards are all correct. This is NOT a code bug — it is a genuine no-edge mechanism.

Well-sampled (6,130 trades) so the negative result is statistically decisive, not noise

Clean gap-safe engineering (1x, 15% notional cap) — it did NOT liquidate, unlike the XRP sibling

avg_trade_return_pct = -0.028% — NEGATIVE per-trade edge, below the 0.15% futures fee floor and below zero; no threshold tuning can flip negative expectancy

profit_factor 0.774 (<1.0), Sharpe -0.87, total_return -94.5%, max_drawdown 94.5% — net loser on every dimension

The iteration-2 z-stop fix FAILED to fix the inverted reward/risk it targeted: avg_win $86.9 vs avg_loss $169.1 (~1.9x), still inverted, and per-trade edge still negative

Every year 2020-2026 is negative and nearly every month is red — no profitable regime exists, so this is a mechanism failure, not a single-window artifact

Falsifies the hypothesis's core premise that the SOL VWAP-reversion recipe ports to ETH; same negative-expectancy VWAP-fade class that has died repeatedly this session

This strategy tried to clone the factory's one surviving price mechanism — SOL intraday VWAP-reversion — onto ETH, an instrument with fee-clearing volatility but no catastrophic-gap history, wrapping it in 1x leverage and a hard 15% notional cap so no gap could liquidate the account. The gap-safety engineering succeeded (it never liquidated), but the edge did not survive the port: across 6,130 trades the per-trade expectancy was negative (-0.028%), profit factor 0.774, Sharpe -0.87, and total return -94.5%, with losses running ~1.9x the average win. Iteration 2 had explicitly added a z-based adverse stop to repair that inverted reward/risk, yet the asymmetry and negative expectancy persisted, and every year from 2020 to 2026 was red — a mechanism failure, not a single-window artifact. The backtest review abandoned it before optimization, judging that tuning entry/exit/stop thresholds — already attempted in iteration 2 — could never convert a structurally negative out-of-sample edge into a real one.

Porting a paper-stage survivor's recipe to a new instrument does not transfer its edge — the VWAP-fade mean-reversion class carries a structural negative-expectancy, inverted-RR signature (small capped reversion wins, large trend-run losses) that has now failed repeatedly across SOL/ETH/XRP/DOGE/BNB/LINK, and adding loss controls addresses the symptom, not the missing directional edge.

abandoned

The pre-optimization backtest review abandoned it: iteration 2's newly added z-based adverse stop failed to fix the inverted reward/risk it targeted, every year 2020-2026 and nearly every month was negative, and no parameter tuning can flip a negative per-trade expectancy the developer had already spent an iteration trying to repair.

A long-short intraday VWAP mean-reversion on ETHUSDT 15m perp bars that faded moderate z-score deviations from a rolling volume-weighted VWAP with trend-gate off, porting the deployed SOL survivor's recipe to ETH behind fixed 1x leverage and a hard 15% notional cap engineered to make liquidation impossible.

On a decisive 6,130-trade sample the edge was confirmed negative — avg_trade_return_pct -0.028% (below zero and the 0.15% fee floor), profit factor 0.774, Sharpe -0.87, total return -94.5%, max drawdown 94.5% — with a 60% win rate undercut by an inverted reward/risk (avg win $86.9 vs avg loss $169.1, ~1.9x). The gap-safe design did work in one sense: unlike the XRP sibling, it never liquidated.