AvaxDailyRsi2OversoldReversionLong

AVAX Daily RSI-2 Deep Oversold Mean Reversion Long

A minimalist long-only mean reversion strategy on AVAXUSDT perpetual futures using daily (1D) bars and the classic Connors RSI-2 indicator (2-period RSI). AVAX is a high-beta smart-contract L1 altcoin that exhibits sharp short-term oversold dislocations followed by mechanical bounces, particularly after sentiment-driven liquidation flushes. The strategy uses a 2-period RSI (extremely sensitive to short-term selling exhaustion) on daily bars to identify deep oversold conditions, then enters long on the next bar's open. The exit is also rule-based and unambiguous: close longs once RSI(2) recovers above a moderate threshold, OR a fixed-bar time stop, OR a hard percentage stop-loss. The strategy has only THREE parameters total (rsi_oversold_threshold, rsi_exit_threshold, max_hold_days) plus a trend filter (200-day SMA). Daily bars + slow signal generation = ~1-3 trades per month, keeping fee drag negligible relative to the 5-15% bounce magnitudes typical of altcoin oversold reversions.

Implements the RSI-2 deep-oversold reversion hypothesis directly. Uses a Wilder 2-period RSI on AVAXUSDT daily bars to detect sharp oversold dislocations, gated by a 200-day SMA uptrend filter so it only buys pullbacks within established uptrends (avoiding falling-knife entries during AVAX bear phases, a key altcoin risk). Exits are fully rule-based and unambiguous: RSI(2)>=60 mean-reversion target, a max_hold_days time stop, or a percentage stop-loss — implemented in should_exit via get_position_pnl_pct. Per the coder anti-frozen-signal guidance, calculate_signal returns a CONTINUOUS bar-varying score (RSI-deviation + price-vs-SMA distance + smoothed momentum, computed from a partial SMA so it never freezes during the 200-bar warmup) while the genuine binary entry decision lives in self._long_trigger, keeping the full trend-filter semantics. Sizing is capital-relative via get_account_equity() and genuinely consumes the declared 2x leverage. Futures venue chosen over spot because the hypothesis specifies perpetual futures and the strategy uses leverage>1. Verified locally: Layer 1 static passes with zero issues; Layer 2 passes all 6 synthetic scenarios with no frozen/NaN warnings. Daily bars + deep-oversold threshold keep trade frequency low (~1-3/month), so round-trip fees (~0.10%) are negligible against typical 5-15% altcoin bounce magnitudes.

Failed deflated Sharpe and every forward gate on attempt 1. DSR=0.1365 (vs 0.95 bar) with the optimized Sharpe 10.04 BELOW the 225-trial expected-max luck bar of 13.72 (is_significant=false, PBO=0.6971>0.5) — after multiple-testing correction the selected config is statistically indistinguishable from best-of-N noise; probabilistic_sharpe 0.9986 vs DSR 0.1365 is the textbook PSR-vs-DSR selection-inflation trap. The walk-forward is is_overfitted=TRUE: avg IS Sharpe 9.45 (windows [9.42, 8.41, 10.53]) collapses to avg OOS 0.0 across ALL THREE windows [0.0, 0.0, 0.0], and the time-ordered holdout FAILED (ratio 0, holdout_sharpe 0.0). Sensitivity FAILED with 5 of 8 cliff parameters (sma_period, max_hold_days, stop_loss_pct, rsi_exit_threshold, rsi_oversold_threshold) — a knife-edge surface where Sharpe survives only at the exact base coordinates and flips negative on ±20% moves. The optimized best_params are degenerate: only 4 trades, profit_factor 0.0, metrics_reliable=FALSE — the apparent 10.04 Sharpe / 1.0 win-rate is a 4-sample artifact, not an edge. Even the base backtest is marginal (Sharpe 1.035 but sharpe_ci_low -2.4989 straddling 0, 32.9% max drawdown, 42 trades over 5.5yr at 2.94% exposure, entire positive contribution from 2021 +57% against negative 2022/2024). Not iterate: the binding failure is all-zero OOS on a 5-cliff surface with no robust region to tune toward — a second best-of-225 sweep re-selects another overfit, sub-significant config. Not revise_hypothesis: per the overfit->abandon rule this is high-IS/zero-OOS overfitting (is_overfitted=TRUE), not a proven mechanism stranded on a dead target — AVAX is a fine instrument and no promoted RSI-2 mean-reversion sibling exists. FAILURE PATTERN: single-asset daily Connors RSI-2 deep-oversold mean reversion overfits a thin 42-trade, 2021-concentrated sample — a modest, CI-straddling base Sharpe (1.035, CI-low -2.50) tuned via best-of-225 yields a knife-edge config (5 cliffs) whose walk-forward OOS is exactly zero in every window and whose optimized backtest collapses to 4 trades; a flattering PSR 0.9986 and a 10-Sharpe full-refit do not rescue an edge that fails deflated Sharpe (0.14), the walk-forward, the holdout, and the sensitivity gate simultaneously.

Long-only Connors RSI-2 deep-oversold mean-reversion on AVAXUSDT.BINANCE USD-M futures, daily bars. Enters long when the 2-period RSI falls below 10 (deep short-term selling exhaustion) while price is above the full 200-day SMA (uptrend filter). Exits when RSI(2) recovers above 60, a 10-day time stop, or a 10% hard stop-loss. Pure OHLCV, three core signal parameters (rsi_oversold_threshold, rsi_exit_threshold, max_hold_days) plus the SMA trend filter. 2x leverage consumed in sizing (~1.0x equity notional at position_pct=0.50).

Functioning backtest with adequate sample: 42 trades over ~5.75 years, sufficient for walk-forward/sensitivity analysis

Genuine modest edge that clears fees: Sharpe 1.04, win_rate 59.5%, profit_factor 1.18, per-trade return ~0.58% (well above the ~0.15% futures round-trip)

Clean data path, no artifacts: sane daily returns, correct sizing (avg_position_pct ~101% = designed 1x notional), positive alpha (0.023) with near-zero beta (0.04)

Very low exposure (2.94%) — selective, brief oversold entries rather than closet beta; capacity to combine or scale

Modest absolute return (25% over 5.75 years, CAGR 5.0%) and mixed annual results (2022 -29%, 2024 -11%) — a thin-ish edge

Wide Sharpe CI (-2.5 to 3.9) and 33% max drawdown reflect the small sample and high-beta altcoin tails; robustness must be confirmed in optimization

42 trades split across 3 walk-forward windows (~14 each) is on the lower side — watch for per-window overfitting

Clean, faithful single-instrument implementation with a decoupled continuous score and full-SMA entry gate (no warmup freeze)

Trades match the hypothesis: 42 long-only entries, deep-oversold RSI(2) with a 200-SMA trend filter, ~1-3/month

Low turnover keeps fee drag modest (commission_pct_of_gross 3.38%)

Deflated Sharpe 0.1365 << 0.95; optimized Sharpe 10.04 BELOW the 225-trial expected-max luck bar of 13.72; is_significant=false, PBO 0.6971

Walk-forward is_overfitted=TRUE: avg IS 9.45 collapses to avg OOS 0.0 across ALL three windows

Holdout FAILED (ratio 0, holdout_sharpe 0.0)

Sensitivity FAILED with 5 cliff parameters — knife-edge surface, Sharpe survives only at exact base values

Optimized config is degenerate: 4 trades, profit_factor 0.0, metrics_reliable=FALSE — the '10 Sharpe' is a sample-size artifact

Even the base backtest is marginal: sharpe_ci_low -2.4989 (CI straddles 0), 32.9% max drawdown, edge concentrated entirely in 2021

This strategy tested the classic idea that high-beta AVAX bounces mechanically after deep short-term oversold dislocations, using a simple RSI(2) signal on daily bars gated by a 200-day trend filter. Its initial backtest looked marginally promising — Sharpe 1.035, 59.5% win rate, positive alpha over 42 trades — enough for the reviewer to greenlight optimization, but the sample was thin and its edge sat almost entirely in 2021 with a Sharpe CI that straddled zero. Optimization then exposed the fragility: in-sample Sharpe of ~9.45 collapsed to exactly 0.0 out-of-sample in every walk-forward window, the holdout and sensitivity gates failed with five cliff parameters, and the best config degenerated to only 4 trades with a deflated Sharpe of 0.1365. The analyst abandoned it on the first iteration, classifying it as textbook high-in-sample/zero-out-of-sample overfitting of a small, era-concentrated dataset with no robust region left to tune toward.

A thin, era-concentrated sample (42 trades, edge mostly from 2021) tuned via a large best-of-N sweep produces a knife-edge, statistically insignificant config; robustness needs a broad stable parameter region and non-zero out-of-sample Sharpe, not a flattering full-refit metric.

The analyst issued an abandon verdict at the post-optimization ANALYZING stage after the strategy failed every forward gate: deflated Sharpe 0.1365 (vs 0.95 bar), is_overfitted=TRUE, holdout failed (ratio 0), PBO 0.6971, and a sensitivity surface with 5 of 8 cliff parameters, with the optimized 10.04 Sharpe falling below the 225-trial expected-max luck bar of 13.72.

A long-only Connors RSI-2 deep-oversold mean-reversion strategy on daily AVAXUSDT perpetual futures, buying short-term selling exhaustion (RSI(2) below threshold) only above a 200-day SMA uptrend filter and exiting on RSI recovery, a time stop, or a hard percentage stop.

The base backtest produced a Sharpe of 1.035 over 42 long trades (~5.75 years) with a 59.5% win rate, 25.4% total return (CAGR 5.0%), and a 32.9% max drawdown, but the Sharpe CI straddled zero (low -2.4989) and the positive contribution was concentrated in 2021. Optimization collapsed: walk-forward went from ~9.45 in-sample Sharpe to exactly 0.0 out-of-sample in all three windows, and the optimized config degenerated to just 4 trades.