SolHyperliquidFifteenMinDailyVwapMeanReversionLong

SOL HYPERLIQUID 15-Minute Daily-Anchored VWAP Mean-Reversion Long-Only with ATR Volatility Filter (Single Asset, 3-Parameter, Short-Horizon, Intraday-Closed)

A LONG-ONLY SHORT-HORIZON intraday mean-reversion strategy on SOL-USD.HYPERLIQUID perpetual using 15-MINUTE bars and OHLCV-only data. Mechanism: classic daily-anchored VWAP mean-reversion at 15m resolution, deliberately modeled on the surviving ETH 1H VWAP Excursion MR pattern (Sharpe 4.70 paper) but transplanted to a NEW VENUE (HYPERLIQUID, 4.9% → 20% target) and a NEW TIMEFRAME BUCKET (short_1m_15m, 2.2% → 10% target) — hitting the TWO most under-represented quotas in the 1113-experiment portfolio simultaneously. On each 15m close, the strategy computes (a) the day's anchored VWAP (sum(close*volume)/sum(volume) since 00:00 UTC, reset at every UTC midnight), (b) the rolling standard deviation of (close - VWAP) across all bars elapsed in the current day, and (c) the z-score = (close - VWAP) / std. Entry requires z-score <= -2.0 (price 2σ below day VWAP, signifying intraday over-extension to the downside) AND a volatility filter: rolling 20-bar ATR-as-pct-of-price >= 0.40% per 15m (filters out grindingly-low-vol periods where the VWAP signal is noise) AND time-of-day < 22:00 UTC (ensures at least 2 hours remain in the UTC day for the position to revert and exit before forced day-end close). Position size 10% of equity, long-only, no leverage requested (HL margin venue but we use 1x). Exit on EARLIEST of: (a) z-score returns to >= 0.0 (close touches or crosses VWAP — primary edge captured), (b) hard daily time stop at 23:45 UTC (close before day-end VWAP reset to avoid stale-anchor exposure), (c) hard PnL stop loss at -1.5% of position notional (cap tail loss). This is explicitly NOT a breakout strategy (the failed mechanism class), NOT a fast-twitch signal (slow daily-anchored VWAP, not RSI(2)), NOT cross-sectional on correlated majors (single asset), NOT drawdown accumulation (no multi-day hold), NOT pair MR with 4-leg fees (single leg, HL fees ~0.09% RT). It directly inherits the SURVIVING ETH-VWAP-MR mechanism class with venue+timeframe transplant for portfolio diversification. SOL is chosen over ETH/BTC because (1) SOL has higher intraday vol than majors so the z=2σ signal triggers more often; (2) SOL/BTC correlation is lower than ETH/BTC, providing portfolio diversification; (3) HYPERLIQUID's lower fees on a fast-trading strategy compound to a meaningful edge advantage over Binance.

Iteration-2 fixes for the two backtest-review issues. (1) DATA COVERAGE: Hyperliquid's candleSnapshot API only exposed ~2.3 months of SOL 15m history (a hard API-lookback/collector limit I cannot extend from strategy code), which is too short for 3-phase walk-forward + a 15-day holdout. Per the feedback's explicit permission, I redirected the identical SOL VWAP-MR mechanism to SOLUSDT.BINANCE, which has multi-year 15m history via Binance Vision, enabling a meaningful WF/holdout. Binance USD-M futures fees (0.10% RT) are the closest low-fee alternative to HL (0.09% RT) and far better than Binance spot (0.20% RT), important given the flagged fee-fragility. This trades off the HL quota but is the only reliable path to the required history; the asset (SOL), timeframe bucket (short_1m_15m), and mechanism are unchanged. (2) WIN/LOSS ASYMMETRY: changed the exit from revert-to-VWAP (z>=0.0, which captured moves too small to clear fees and the stop) to a VWAP-overshoot target (z>=0.5), making winners ~25% larger, and tightened the hard stop from 1.5% to 1.2% so losers stay smaller than the overshoot winners — directly targeting the PF 0.73 / avg_loss>>avg_win profile. All signal computation, ATR/time gates, and sizing are unchanged to preserve every previously-passing verification layer.

Consistent structural money-loser with no edge: negative total return (-44.2%) and NEGATIVE returns in all 7 years (2020->2026), including the 2021 and 2024 bull markets, so this is edge-absence, not a regime or sizing problem. Risk-adjusted metrics are decisively negative (Sharpe -2.43, Sortino -3.33, profit_factor 0.82, expectancy -16.8/trade, PSR 0.0, Sharpe CI entirely below 0). Fees dominate what little signal exists: commission_pct_of_gross=11.26% over 2,529 trades, with avg_win 183 barely above avg_loss 167 at a losing 42.8% win rate. This is already iteration 2 and exposes only 3 tunable params (entry_z, atr_min_pct, position_pct), none of which can convert an all-years-negative expectancy into a positive one. The ETH-1H VWAP mean-reversion pattern did not survive the transplant to SOL 15m — the 4x-shorter timeframe multiplied trade frequency and handed the edge to trading costs. No optimization can fix an edge that is below trading costs; matches the avg_trade_return-below-costs abandon rule with no articulable improvement path.

Long-only intraday daily-anchored VWAP mean-reversion on SOLUSDT 15-minute bars. Each 15m close computes the day's volume-weighted VWAP (reset at UTC midnight), the std of (close - VWAP) over the day's elapsed bars, and z = (close - VWAP)/std. Enters LONG when z <= -2.0 (2σ below day VWAP) with a 20-bar ATR%>=0.40% volatility filter and time-of-day < 22:00 UTC. Exits at the earliest of z >= +0.5 (VWAP-overshoot target so winners are larger), 23:45 UTC day-end close, or a -1.2% hard PnL stop. 10% of equity, 1x, OHLCV-only.

Clean backtest: no liquidation, no sizing blowup, sensible ~11% avg position, fires often (2,529 trades) so sparsity is not the problem

benchmark_meaningful=true and the mechanism is well-specified

Loses money in EVERY year of the sample (2020 through 2026), including the 2021 and 2024 bull markets — the signal has no edge, not a regime issue

Deeply negative risk-adjusted metrics: Sharpe -2.43, Sortino -3.33, profit_factor 0.82, expectancy -16.8/trade, total_return -44.2%, probabilistic_sharpe 0.0, Sharpe CI entirely below zero (-3.23 to -1.54)

Fee drag dominates: commission_pct_of_gross=11.26%, $22.4k commissions over the run; the tiny per-trade edge (avg_win 183 vs avg_loss 167 at 42.8% win rate) is negative before fees and buried after

Already iteration 2 (HL->Binance data redirect) with only 3 tunable params (entry_z, atr_min_pct, position_pct) — none can flip an all-years-negative expectancy positive

The 'surviving ETH 1H VWAP-MR' mechanism did not survive transplant to SOL 15m: the 4x-shorter timeframe multiplied trade count and handed the edge to fees

Do NOT optimize — the backtest only covers ~2.3 months of HL SOL 15m data (2026-03-16 to 2026-05-24, 29 active days, 46 trades), which is far too short for the 3-phase optimization (walk-forward needs 3 IS/OOS windows and the 15-day holdout would consume half the data; the result would be fit to noise). TWO things to fix before re-submitting: (1) PRIMARY — DATA COVERAGE: HYPERLIQUID SOL perp has traded since ~2023, so 15m history should extend well beyond 2.3 months. Investigate why only ~2 months loaded (likely a limited on-demand collection window or a 15m catalog gap) and extend the HL SOL 15m history to at least ~2 years so a meaningful walk-forward + holdout can run. If more HL SOL 15m history genuinely cannot be collected, redirect the same mechanism to a target with multi-year 15m history (note this trades off the HL quota goal). (2) SECONDARY — FIX THE WIN/LOSS ASYMMETRY: even on the short window the profile is fee-fragile (PF 0.73, avg_loss $164 vs avg_win $64, commission 21.5% of gross). The z>=0 revert-to-VWAP exit captures moves too small to clear the HL round-trip + the -1.5% stop. Consider a VWAP-overshoot target (exit at z>=+0.3 to +0.5 rather than z>=0) so winners are larger than the stop-out losers, and/or tighten the stop relative to the target. Re-backtest on the extended data; if the longer history shows positive expectancy with the fixed exits, it's worth optimizing.

This strategy transplanted the surviving ETH 1H VWAP mean-reversion pattern to SOL at a 15-minute resolution, buying intraday dips ~2σ below the daily-anchored VWAP with ATR and time-of-day filters, and was redirected from Hyperliquid to Binance in iteration 2 after Hyperliquid offered too little 15m history for walk-forward testing. The backtest was clean mechanically (no liquidation, ~11% average position, 2,529 trades) but a decisive financial failure: -44.2% total return, Sharpe -2.43, profit factor 0.82, and losses in all seven years including two bull markets, with commissions eating 11.26% of gross. The analyst abandoned it at the backtest-review gate, concluding the edge was absent rather than merely mis-tuned and that no optimization over its three parameters could rescue an expectancy that was negative even before fees. The core takeaway: the slower-timeframe VWAP edge did not survive the shift to a 4x-faster timeframe, where trade frequency handed the strategy's economics to trading costs.

A mean-reversion edge that survives at a slower timeframe (the ETH 1H VWAP pattern) does not automatically transplant to a 4x-faster 15m timeframe, where the multiplied trade count hands any thin edge to trading costs — and three tunable parameters cannot flip an all-years-negative expectancy positive.

It was abandoned at the pre-optimization backtest-review gate (verdict: abandon) as a consistent, edge-absent money-loser whose tiny per-trade signal was already negative before fees and buried after; optimization, analyst, and risk-review stages were never reached.

A long-only, single-asset intraday mean-reversion strategy on SOL 15-minute bars that buys when price falls ~2σ below the daily-anchored VWAP (gated by a 20-bar ATR volatility filter and a time-of-day cutoff) and exits on a VWAP overshoot, a day-end close, or a hard PnL stop.

Over ~1330 data days it traded 2,529 times but lost money structurally: total return -44.2%, Sharpe -2.43, Sortino -3.33, profit factor 0.82, expectancy -16.8 per trade, and a 42.8% win rate with avg_win (183) barely above avg_loss (167). Commissions consumed 11.26% of gross (~$22.4k), and it posted negative returns in every year from 2020 through 2026, including the 2021 and 2024 bull markets.