Strategies

TRX Dual-Timeframe Momentum Confluence, Long-Short (Single-Instrument TRXUSDT.BINANCE USD-M, 4H Primary + 1D Confirm, Flat on Disagreement, ATR-Trailing Exit, 3-Parameter) — Applying the Factory's One Proven Survivor to a Structurally Low-Volatility Persistent Trender, Not Another Whippy High-Beta L1

Outcome: Abandoned

TrxDualTimeframeMomentumConfluenceLS4H

Outcome Summary

This strategy re-used the factory's only paper-stage-reaching mechanism — 4H+daily momentum confluence, long-short, flat on disagreement — and applied it to TRXUSDT on the thesis that TRX's low-volatility USDT-rail drift would trend like the BTC/ETH/ADA survivors instead of whipsawing like the failed L1 symbol-swaps. On a decisive 441-trade, 6.6-year backtest it produced profit_factor 0.94 and a negative -$78.28/trade dollar expectancy; the +34% headline was pure unrealized open position and the 1.34 Sharpe an MTM artifact of ~full-notional sizing that ran the account into a 79% drawdown, while 202% impact cost and $24k capacity meant any real edge existed only at toy scale. The analyst abandoned it at the backtest-review gate for negative expectancy, judging that three parameters and walk-forward optimization could not lift an edge that was net-negative before tuning. The instrument thesis was falsified: TRX behaved like the failed clones, not the winners, losing money in the calm regime and in every year except a 2024 artifact spike.

Strategy report

Hypothesis

A LONG-SHORT, single-instrument, pure-OHLCV momentum-confluence trend follower on TRXUSDT.BINANCE (USD-M perpetual). Mechanism is IDENTICAL to the three strategies that actually reached paper_stage (BTC 4H+1D confluence Sharpe 1.99, ADA Sharpe 1.60, ETH TSMOM 1.06): take a directional position ONLY when a FAST (4-hour) trend signal AND a SLOW (daily) trend signal AGREE in sign; stand flat when they disagree. This is deliberately NOT a new mechanism — the meta-learner has shown (L92) that bolting regime/efficiency/volatility gates onto momentum does not create edge, and (L88, L77, L45) that baskets, TSMOM books, and cross-sectional rotations are dead. The ONLY thing that survives here is slow directional confluence on a CLEAN-TRENDING large cap. The novel, falsifiable claim is about the INSTRUMENT, not the recipe: every failed symbol-swap of this family (XRP, LINK, SOL, AVAX, NEAR, DOT — all overfit/fee_edge) is a high-beta, mean-reverting, whipsaw-prone L1 whose 4H trend flips constantly and manufactures losing round-trips; the three SURVIVORS (BTC, ETH, ADA) are large, liquid, persistently-trending names with lower relative noise. TRX has a structurally distinct return process from the failed L1s: demand is anchored to Tron's dominant USDT-settlement flow (the largest stablecoin transfer rail), giving it a persistent low-volatility drift and far fewer whipsaws than the high-beta alts — a regime profile closer to the BTC/ETH/ADA winners than to the failures. TRX has never been tried in the corpus (74 of 767 collected instruments were ever used) and has multi-year Binance USD-M history (listed 2020) for full statistical power. 3 parameters only, to resist the overfitting that killed the symbol-swap clones.

This is a deliberate, mechanism-identical port of the only momentum family that has reached paper_stage in this factory (BTC 4H+1D confluence Sharpe 1.99, ADA, ETH TSMOM): fast trend AND slow confirm must agree, flat on disagreement, ATR-trailing exit, three parameters. Per the hypothesis the novelty is the INSTRUMENT, not the recipe, so no regime/efficiency/volatility gate is bolted on (L92) and no basket/rotation structure is used (L88/L77/L45). Both legs are computed on the single 4H close series at two horizons rather than subscribing a separate 1-DAY leg: the slow leg at 150 4H bars IS the daily-scale confirm in economic terms, it matches the proven implementation exactly, and it avoids the multi-leg data-alignment/injection failure mode that has repeatedly cost iterations (L86/L35). Venue is Binance USD-M futures because the strategy takes short positions; leverage stays 1.0 (notional capped at equity) because the edge is directional, not leverage-dependent, and L15 flags compounding-sizing blowups. Parameter defaults were chosen from the real TRXUSDT 4H history (14,554 bars, 2020-01 to 2026-09) by locating a broad plateau rather than a peak: across fast in {24,30,36}, slow in {120,150,180} and trail in {8,10,12} the mechanism is uniformly net-positive after 0.10% round-trip costs, with ~467 trades (~70/yr, ample statistical power), average gross per-trade return ~0.36% (well above the 0.15% futures fee floor) and a last-30%-of-history out-of-sample slice that is STRONGER than in-sample (~0.70% per trade) - the opposite of the overfit signature that killed the XRP/LINK/SOL/AVAX symbol swaps. Note the original BTC defaults (12/42/5.0) are NOT viable on TRX (~0.0002% per trade, net negative): TRX needs slower horizons, which is consistent with the hypothesis's claim that it is a low-volatility persistent trender rather than a whippy high-beta L1. No parameter is clamped in code, so no _param_bounds are declared and the optimizer's mechanical 0.5-1.5x search stays inside the tested plateau.

Negative expectancy / net-losing signal on a decisive 441-trade, 6.6-year sample: profit_factor 0.9398 (<1 after costs), dollar expectancy -$78.28/trade (avg_trade_pnl_usd negative). The positive avg_trade_return_pct (0.387%) is a mean-of-percents artifact inflated by a few blow-off winners (largest_win $94k) while the account bleeds — the honest dollar economics are negative, so optimization cannot lift an edge that is net-negative before it can even be tuned. Two structural killers on top: (1) max_drawdown 79.4% (ci_low 93%), recovery_factor -1.0, calmar -1.0, and impossible daily prints (+410%, +457%, +292%, -426% on 2023-09-01) show ~full-notional sizing (avg_position_pct 352%) riding the equity curve into repeated near-blowups — the Sharpe 1.34 is an MTM artifact and end_unrealized_pct 68.9% means the +34% headline is entirely open-position unrealized; (2) impact_cost_pct 202% with capacity_usd only $24,469 means the nominal edge survives only at toy scale and is erased at deployable size (QA's capacity/impact concern confirmed and worse on full history). This is the single-name OHLCV dual-TF momentum-confluence clone family (L82/L93); the hypothesis's own thesis that TRX would behave like the BTC/ETH/ADA survivors is falsified — TRX sits far below the survivor tier (PF>=2.0, Sharpe>=1.5, DD<20%), losing in the calm regime (-62.9%) and in every year except the 2024 artifact spike. Three parameters cannot rescue a PF-0.94, negative-expectancy, 79%-DD signal; not worth 2 hours of walk-forward. abandon_class=negative_expectancy.

Implementation

Long/short dual-horizon momentum confluence on TRXUSDT.BINANCE (USD-M perpetual), 4-HOUR bars, pure OHLCV, single instrument. Two momentum horizons are computed on the same 4H close series: a FAST leg (30 bars, ~5 days) and a SLOW daily-scale confirm leg (150 bars, ~25 days). A directional position is opened ONLY when both legs agree in sign (LONG if both positive, SHORT if both negative) and the strategy stands FLAT whenever they disagree. Exits are (a) the slow confirm leg flipping against the position, and (b) an ATR trailing stop at trail_atr_mult x ATR from the best price since entry, so confirmed trends run while chop is cut. Sizing is capital-relative: 1.5% of equity risked against a 1.5 x ATR initial stop distance, hard-capped at 1.0 x equity notional so losses can never compound into position size. Three edge tunables only (fast_lookback, slow_lookback, trail_atr_mult); leverage 1.0.

Verification Results

None required for correctness. If exact parity with a 1D confirm is ever desired, subscribe a 1-DAY extra_bar_type; current approach is a reasonable and defensible substitute.

The hypothesis title states '4H Primary + 1D Confirm', but the slow confirm leg is a 150-bar lookback computed on the SAME 4H close series (~25 days), not a separate subscribed 1-DAY bar. This is economically a daily-scale trend confirm and the developer justifies it as matching the proven implementation while avoiding multi-leg data-injection failures. Core edge (slow directional confluence, flat on disagreement) is faithfully present, so this is acceptable, but note the slow leg is an approximation of a true daily bar rather than a literal 1D series.

Confirm the base template actually consults an override of min_bars_required; if not, the self-guard already makes warmup safe. No functional impact observed.

min_bars_required() shadows the framework method (static-analysis base_shadow warning); the base may rely on its own warmup implementation. Low risk here because calculate_signal self-guards (only computes momentum when len(closes) > slow_lookback+1), so no signal fires before sufficient history regardless.

Sandbox shows high market-impact fragility: impact_cost_pct 70.5% of gross with capacity only ~$201k, driven by ~100% equity notional per trade (avg_position_pct 99.8%) on the lower-liquidity TRXUSDT — the 0.337% per-trade edge survives only at toy scale and erodes at size. Also a very deep calm-regime drawdown (calm tercile return -98%, overall max_drawdown ~78.7%). Not a correctness defect (sizing is standard capital-relative with a 1.0x notional cap), but the analyst should stress capacity/impact and calm-regime behavior on full history before optimizing.

Backtest Review

Sharpe: 1.34
Total return: 34.38%
Max drawdown: 79.39%
Trades: 441
Win rate: 25.6%
Profit factor: 0.94

Decisive sample: 441 trades over 2426 days (6.6 years), full Binance USD-M history, balanced 214 long / 227 short

Code faithfully implements the dual-TF confluence hypothesis (441 signaled = 441 submitted, flat-on-disagreement, ATR trailing exit) — this is a fair test, not a broken mechanism

profit_factor 0.9398 (<1) — net losing after costs; expectancy / avg_trade_pnl_usd = -$78.28 per trade (NEGATIVE dollar expectancy)

The +0.387% avg_trade_return_pct is a mean-of-percents artifact inflated by a handful of blow-off winners (largest_win $94k) while the account bleeds — PF 0.94 confirms net-negative economics

max_drawdown 79.4% (max_drawdown_ci_low 93%), recovery_factor -1.0, calmar -1.0 — far past the >50% hard-abandon line

impact_cost_pct 202% (impact > 2x gross PnL) and capacity_usd only $24,469 — the nominal edge exists only at toy scale, erased at any deployable size (QA concern confirmed and worse on full history)

avg_position_pct 352% with impossible daily prints (+410%, +457%, +292%, -426%) = ~full-notional sizing riding the curve into repeated near-blowups; Sharpe 1.34 is an MTM artifact of that curve

end_unrealized_pct 68.9% — the +34% headline is entirely open-position unrealized; realized track record is negative

Loses in calm regime (-62.9%) and in every calendar year except the 2024 artifact spike (+1967%); win_rate 0.256

Outcome Summary

Transplanting a proven mechanism onto an untried instrument is not itself an edge — TRX's return process fell far below the survivor tier, and dollar expectancy and profit factor must be checked before trusting an avg_trade_return_pct that a handful of blow-off winners can inflate.

The analyst issued an 'abandon' verdict at the pre-optimization backtest-review gate (abandon_class=negative_expectancy): the signal is net-losing after costs, its positive avg_trade_return_pct (0.387%) is a mean-of-percents artifact inflated by a few blow-off winners (largest_win $94k), and three parameters plus optimization cannot rescue a PF-0.94, negative-expectancy, 79%-drawdown edge — so it never reached optimization, analysis, or risk review.

It ported the factory's one proven survivor mechanism — dual-timeframe (4H fast + daily-scale confirm) momentum confluence, long-short and flat on disagreement with an ATR-trailing exit — onto TRXUSDT.BINANCE USD-M, betting TRX's low-volatility, USDT-settlement-anchored drift would trend like the BTC/ETH/ADA winners rather than whipsaw like the failed high-beta L1 clones.

Over a decisive 441-trade, 6.6-year sample (2020-2026, balanced 214 long / 227 short), it posted profit_factor 0.94 (<1) with negative dollar expectancy of -$78.28/trade and a 25.6% win rate. The headline +34% return was entirely open-position unrealized (end_unrealized_pct 68.9%), the reported Sharpe 1.34 was an MTM artifact of ~full-notional sizing (avg_position_pct 352%), and max drawdown hit 79.4% (ci_low 93%) with impact_cost_pct 202% against capacity of only $24,469.

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