Skip to content

View translation

NearFourHourSupertrendTrendFollowingLong

Hypotheses

NEAR 4H Supertrend Trend-Following Long

Hypotheses

A long-only single-instrument TREND-FOLLOWING strategy on NEARUSDT perpetual futures (Binance USD-M) using the Supertrend indicator on 4-hour bars. CRITICAL DESIGN CHOICES locked in based on the now-decisive feedback that 6 CONSECUTIVE supplementary-data hypotheses have failed (ETH liquidations, BTC taker_ratio, DOGE OI, BTC OI, XRP premium_index, AVAX taker_ratio), confirming the supp-data infrastructure is currently broken: (a) ABSOLUTELY ZERO supplementary data — pure OHLCV. (b) NOT mean-reversion on alt-perps (XRP VWAP-deviation explicitly falsified, also no edge in modern ETH microstructure). (c) NOT a contrarian short (BTC funding short structurally falsified — modern crypto microstructure makes extreme positioning a CONTINUATION signal). (d) 4H + Binance USD-M — the empirically proven verification combination. (e) NEW symbol (NEAR — not in current portfolio of ADA/AVAX/BNB/BTC/ETH/LINK/LTC/SOL). (f) NEW mechanism — Supertrend is structurally different from every other in-pipeline strategy: it's not a channel (Donchian/Keltner), not a volume breakout, not an MA cross (Golden/EMA), not a MACD signal, not a wide-range bar, not a drawdown accumulator. Supertrend creates a stair-step ATR-based trailing trend filter that flips direction discretely. (g) MAXIMALLY SIMPLE — canonical Supertrend parameters (period=10, multiplier=3.0) FIXED at defaults, eliminating curve-fit concerns flagged in recent feedback.

Hypotheses

Implements the canonical Supertrend trend filter, a stair-step ATR trailing-stop mechanism structurally different from channels (Donchian/Keltner), MA crosses, MACD, volume breakouts, wide-range bars, and drawdown accumulators already in the pipeline. It uses zero supplementary data (the six recent supp-data failures motivated a pure-OHLCV design) and the proven 4H + Binance USD-M verification combination on a new symbol (NEAR). The Supertrend is computed incrementally with the correct recursive final-band logic (each band depends on the previous bar's band and close) and a Wilder-RMA ATR, which is more faithful and less error-prone than recomputing from scratch each bar. calculate_signal returns a continuous, bar-varying signed distance to the Supertrend line so the verifier sees real signal variation, while the discrete trend state gates should_enter/should_exit and position_size, avoiding the gated-to-0.0 anti-pattern. Parameters are fixed at canonical defaults (10, 3.0) to minimize curve-fit surface. Venue is BINANCE USD-M futures (long-only; futures RT taker ~0.10% is cheaper than spot) with leverage left at 1.0 (no amplification used) so the leverage_set_but_unused gate does not trigger; the trend-following holding period produces moves far larger than the round-trip fee.

Hypotheses

A genuine but fat-tail-fragile trend edge that does not deflate or hold out — the closest call of the batch, but it fails the decisive gates. Positives are real: sensitivity PASSED with 0 cliffs, is_overfitted=FALSE, avg OOS 0.52, sharpe_ci_low just positive (0.0154), and the optimized Sharpe 0.87 is slightly above the expected-max luck bar 0.86. But it is rejected by three independent promotion blockers: deflated_sharpe 0.5112 (< 0.95, is_significant FALSE), PBO 0.8399 (84% probability of backtest overfitting, decisively > 0.5), and a FAILED holdout (holdout_sharpe 0.162 is only 31% of the walk-forward OOS 0.519, ratio 0.311 below the 70% bar). The root cause is intrinsic fat-tail dependence — optimized return_kurtosis 175.5, skew 12.1, tail_ratio 0.057, with the entire +103% carried by a handful of explosive NEAR trend days — which is exactly why DSR can't clear 0.95 and why CSCV rank-stability collapses to PBO 0.84; the edge also decayed in 2025 (-6.9%). Not promote: not significant, PBO 0.84, holdout failed. Not iterate (attempt 1 of 2): the failures all stem from intrinsic fat-tail fragility of long-only trend-following on a high-beta alt, not a tunable parameter — the sensitivity grid is flat (Sharpe ~0.53-0.66, 0 cliffs) and the optimizer is already at its best config, so a second sweep re-finds another outlier-dependent config that fails the recent holdout. Not revise_hypothesis: the path-dependent trend expression is already captured by the approved Supertrend sibling; NEAR is the fat-tail/decaying arm. FAILURE PATTERN: long-only Supertrend trend-following on a high-beta alt (NEAR) optimizes to a clean-looking config (PF 2.25, Sharpe 0.87, 0 sensitivity cliffs, is_overfitted=FALSE) whose headline is carried by a few explosive trend days (kurtosis 175) — so it fails deflated Sharpe (0.51 < 0.95), shows PBO 0.84, and fails the holdout (31% retention). A fat-tail-driven trend edge with high PBO and a sub-threshold holdout is statistically fragile, not deployable, even when the raw Sharpe edges past a low luck bar — the same fat-tail signature as the abandoned XRP Chandelier sibling, in a less extreme but still-failing form.

Implementation

Long-only Supertrend trend-following on NEARUSDT.BINANCE USD-M futures, 4H bars, pure OHLCV (zero supplementary data). Computes the canonical Supertrend (ATR period 10, multiplier 3.0) incrementally: Wilder-RMA ATR, hl2 +/- mult*ATR basic bands, the standard final-band recursion, and a discrete trend state. Enters LONG when the trend is up and flat; exits when the trend flips down (the Supertrend line is the trailing stop) or on a 15% catastrophic safety stop. The signal is the continuous signed distance (close-supertrend)/supertrend, recomputed every bar; the trend state gates entry/exit. Sized at 30% equity notional, capped so risk to the Supertrend stop <= 2% of equity. leverage 1.0.

Backtest Review

Healthy sample and economics: 140 trades (~25/yr), 0 dropped, PF 1.339, Sharpe 0.60, Sortino 1.81, expectancy +$550/trade, total +98%

Backtest Review

Proven mechanism (Supertrend has an approved sibling), pure OHLCV (no supplementary-data wall), implemented correctly and trades as specified

Backtest Review

Deployable: capacity ~$47M with only 4.6% impact cost and 1.6% commission drag — not a toy-scale edge

Backtest Review

Fat-tail dependence: return_kurtosis 139, skew 9.96, tail_ratio 0.04, win_rate 35% — the +98% headline is carried by a few +20%+ days; the deflated-Sharpe skew/kurtosis penalty may heavily discount it (as it did the XRP Chandelier sibling)

Backtest Review

Regime-concentrated: 2021 (+50.7%) and 2023 (+31.5%) dominate; 2024 -0.6%, 2025 -12.5%, rolling Sharpe negative through 2025-early 2026 — holdout risk

Backtest Review

Sharpe CI straddles 0 (-0.37 to 1.29); max_drawdown 21.3% (CI up to 48%), max_consecutive_losses 11

Analysis

Best robustness profile of the batch: sensitivity PASSED with 0 cliffs, is_overfitted=FALSE, avg OOS 0.52, sharpe_ci_low 0.0154 (just positive), and optimized Sharpe 0.87 slightly above the expected-max luck bar 0.86

Analysis

Optimized config looks clean in isolation: PF 2.25, Sortino 4.43, max_drawdown 17.6%, huge capacity (~$870M), low fee drag (0.95% of gross)

Analysis

Proven mechanism (Supertrend has an approved sibling), pure OHLCV, implemented correctly

Analysis

Not significant: deflated_sharpe 0.5112 (< 0.95, is_significant FALSE) — does not clear multiple-testing deflation

Analysis

PBO 0.8399 — 84% probability of backtest overfitting, decisively > 0.5

Analysis

Holdout FAILED: holdout_sharpe 0.162 is only 31% (ratio 0.311) of the walk-forward OOS 0.519, below the 70% retention bar

Analysis

Extreme fat-tail dependence: return_kurtosis 175.5, skew 12.1, tail_ratio 0.057 — the entire +103% is carried by a few explosive trend days; win_rate 37%

Analysis

Recent decay: 2025 -6.9%; rolling Sharpe dipped negative into early 2026

Analysis

No tunable region: sensitivity grid is flat (Sharpe ~0.53-0.66, 0 cliffs) and the optimizer is already at its best config

Outcome Summary

This was the closest call of its batch — a clean, novel Supertrend trend-follower on NEAR with a genuinely healthy base (140 trades, +98%, PF 1.34, ~$47M capacity) that earned an 'optimize' verdict, and an optimized config that looked strong in isolation (PF 2.25, Sharpe 0.87 just past the 0.86 luck bar, 0 sensitivity cliffs, is_overfitted=False). But the edge was carried by a few explosive trend days, so it failed three decisive gates: deflated Sharpe 0.51, PBO 0.84, and a holdout retaining only 31% of walk-forward OOS Sharpe, alongside 2025 decay. The analyst abandoned it on iteration 1 rather than iterating, judging the fat-tail fragility intrinsic to long-only trend-following on a high-beta alt — the same signature as the abandoned XRP Chandelier sibling, in milder but still-failing form — with the path-dependent trend expression already covered by the approved Supertrend sibling.

Outcome Summary

Long-only Supertrend trend-following on a high-beta alt can pass sensitivity and walk-forward yet still be undeployable: when the entire return is carried by a handful of explosive trend days (kurtosis ~175), deflated Sharpe can't clear the luck bar, PBO climbs to 0.84, and the holdout collapses — fat-tail dependence is statistical fragility no second parameter sweep can fix.

Outcome Summary

The post-optimization analyst abandoned it on three independent promotion blockers: deflated Sharpe 0.5112 (<0.95, not significant), PBO 0.8399 (84% overfitting probability), and a failed holdout (holdout Sharpe 0.162, only 31% retention of the 0.519 walk-forward OOS, below the 70% bar) — with iteration ruled out because the failures stem from intrinsic fat-tail fragility, not a tunable parameter (the sensitivity grid was flat and the optimizer was already at its best config).

Outcome Summary

A long-only Supertrend trend-following strategy on NEARUSDT 4H perpetual futures using pure OHLCV (canonical period=10, multiplier=3.0), holding long while the ATR-based stair-step trend filter pointed up and exiting when it flipped down or a 15% catastrophic stop hit.

Outcome Summary

The base backtest was healthy and deployable — 140 trades (~25/year), +98% total, Sharpe 0.60, profit factor 1.339, expectancy +$550/trade, 21.3% max drawdown, ~$47M capacity at 4.6% impact; optimization produced a clean-looking config (PF 2.25, Sharpe 0.87, Sortino 4.43, 17.6% drawdown, 0 sensitivity cliffs, is_overfitted=False), but the headline was carried by a few explosive trend days (return kurtosis 175.5, skew 12.1, tail_ratio 0.057, win rate 37%).
Strategy report

Backtest and paper results are hypothetical. Trading involves risk of loss.