Strategies

FIL Daily Multi-Week Trend Continuation Long with 200-SMA Regime Filter

Outcome: Abandoned

FilDailyMultiWeekTrendContinuationLong

Outcome Summary

This strategy replicated the promoted LINK and ATOM daily multi-week trend-continuation mechanism on Filecoin (FIL) for portfolio diversification, using a 200-SMA regime filter and a 30/80 EMA momentum stack calibrated for a handful of trend-riding entries per year. Instead, the backtest lost roughly two-thirds of capital (-68.8% return, Sharpe -0.519, 79% max drawdown) across 175 trades and betrayed its own design, firing ~30 entries a year with ~1-day average holds as the logic whipsawed in FIL's choppy, largely bearish regimes. The reviewer identified the binding problem as structural — a long-only trend-follower applied to an asset that fell ~98% from its 2021 high has no uptrend edge for optimization to recover — and abandoned it at the backtest-review gate after five iterations, before spending any time on optimization or analysis.

Strategy report

Hypothesis

A long-only single-instrument trend-continuation strategy on FILUSDT perpetual futures using daily bars and OHLCV-only data. Architecturally mirrors the proven LINK 1D Multi-Week Trend Continuation (Sharpe 2.55, promoting) and ATOM Daily Multi-Week Trend Continuation (already pending), applied to FIL (Filecoin) for genuine portfolio diversification across uncorrelated narrative drivers. FIL is fundamentally distinct from every existing pipeline strategy's underlying narrative: BTC (institutional store-of-value), ETH/SOL/ATOM (smart-contract platforms), BNB (exchange utility), LINK (oracle middleware), TRX (stablecoin transport), XRP/payments. FIL is the largest decentralized storage token, with cycle dynamics driven by: (a) Filecoin Plus subsidy program growth and Storage Provider deal-making activity, (b) FVM (Filecoin Virtual Machine) launch and DeFi-on-storage narratives, (c) AI training data storage demand (post-2024 GPT-style narrative), (d) USDC/RWA-on-Filecoin integration progress. These drivers produce multi-month trend phases that are largely orthogonal to BTC institutional flows or ETH DeFi activity — giving genuine portfolio-level diversification value beyond just adding 'another altcoin' to the basket. Single-dominant-filter design (200-day SMA regime + 30/80 EMA stack momentum), explicitly NOT a multi-condition AND-gate (the recurring failure mode of AVAX Calm-Regime / DOT Pullback / SOL TOM, all abandoned for sparse-signal reasons). Calibrated for 5-12 entries/year (the empirically validated success zone of LINK Multi-Week Trend). FIL has 5+ years of clean Binance USD-M perp data covering the 2021 bull (high of $237), 2022-2023 bear (low of $2.84), and 2024 partial recovery — providing rich walk-forward validation across multiple regimes.

Iteration 5 fix for the Layer-2 'steady_downtrend: frozen signal, all 296 signals identical (value=0.0)' failure. Root cause: the previous calculate_signal bailed the whole window to a bare `return 0.0` whenever `np.any(closes <= 0)` — and the production synthetic steady_downtrend generator declines steeply enough to drive prices to/through zero, so that guard fired on every post-warmup bar, freezing the series at exactly 0.0 (matching the reported value). Uptrend/flat never cross zero, which is why only downtrend froze. Smallest fix: introduce a difference-based (not price-ratio) per-bar momentum backbone `_mom_score` that stays finite even when prices go <= 0, and replace every former bare-`0.0` guard path (warmup, non-positive prices, degenerate sma/ema) with `return mom_score` so the signal always varies bar-to-bar; the only constant return is the single first bar (n < 2). Verified locally: all 6 real scenarios yield 296 unique signals, AND a simulated zero-crossing downtrend (min close −100) yields 296 unique finite signals with no freeze. All earlier-passing logic is preserved unchanged — imports, class structure, EMA/SMA math, the regime+momentum _long_trigger entry gate, exits (trailing stop / EMA break / regime break), and leverage-aware sizing — so no earlier layer regresses and real-data trading behavior is identical (real prices are always > 0, so the full model path runs exactly as before).

Catastrophic no-edge result plus a mechanism mismatch, on an ample 175-trade sample. The base backtest loses money decisively (total_return -68.8%, profit_factor 0.649, Sharpe -0.519, negative expectancy -$430/trade, max_drawdown 79.0%). Critically, the trades do NOT implement the hypothesis: the 'multi-week trend continuation' calibrated for '5-12 entries/year' with holds that 'let trends run' actually produced ~30 entries/year with a ~1-day average holding period — the entry (close>SMA200 AND EMA30>EMA80) and exits (price<SMA200 / EMA-stack break) whipsaw on consecutive bars in FIL's choppy/bearish regimes rather than riding multi-week trends. The root cause is a premise problem optimization cannot fix: a long-only trend-follower applied to a structurally bearish asset (FIL fell ~98% from its 2021 high of $237 to $2.84), so there is no uptrend edge to tune toward and the 175-trade sample already establishes a decisive no-edge result. The promoted LINK/ATOM siblings worked because those instruments trended; porting the same architecture to FIL does not inherit promotability. Not worth 2 hours of optimization — the base fails to implement its own multi-week premise and loses two-thirds of capital with a 79% drawdown.

Implementation

Long-only FILUSDT daily-bar futures trend-continuation. Enters long when the 200-day SMA regime filter (close > SMA) AND the 30/80 EMA momentum stack (EMA_fast > EMA_slow) are both aligned up — a single-dominant-filter design (not a sparse multi-condition AND-gate). calculate_signal() returns a continuous, always-varying trend score (difference-based per-bar momentum backbone plus EMA-spread / price-vs-SMA / short-momentum blend); the binary regime+momentum decision is stored in _long_trigger. Exits let trends run: 8% trailing stop from the in-trade peak close, EMA-stack break, or a drop below the 200-SMA. Leverage 2.0 is genuinely consumed in sizing (~1.2x equity notional). Calibrated for ~5-12 entries/year (LINK Multi-Week Trend success zone).

Verification Results

Verification failed (Layer 2 — synthetic scenarios): Parameters used: ['ema_fast', 'ema_slow', 'trail_pct', 'min_notional', 'position_pct', 'regime_period'] Check that __init__ sets all attributes from self.parameters.get(). - steady_downtrend: Frozen signal: all 296 signals are identical (value=0.0). calculate_signal() likely has a bug — the signal never varies.

Backtest Review

Sharpe: -0.52
Total return: -68.80%
Max drawdown: 79.00%
Trades: 175
Win rate: 45.7%
Profit factor: 0.65

Clean infrastructure: single-asset FIL 1D OHLCV-only, long-only, 175 entries all submitted with zero drops; ample sample size (not a sparsity case).

Catastrophic money-loser with no edge: total_return -68.8%, profit_factor 0.649, Sharpe -0.519, negative expectancy (-$430/trade), max_drawdown 79.0% (CI to 97.7%).

Mechanism mismatch: hypothesis promises 5-12 entries/year with multi-week holds, but the backtest produced ~30 entries/year with a ~1-day average hold — the trend-continuation logic is whipsawing, not letting trends run.

Long-only applied to a structurally bearish asset (FIL fell ~98% from its 2021 high) — annual returns negative in 2021 (-48.5%), 2023 (-56.2%), 2025 (-2.8%); the only positive year is 2024.

Negative alpha (-0.08) and information_ratio (-0.05); 9 consecutive losses at worst.

Analysis

The strategy does not implement its stated multi-week continuation mechanism: it fires 175 entries with ~1-day average holds instead of the calibrated 5-12 entries/year with multi-week holds. Root cause is an exit/re-entry churn loop. Fixes: (1) Remove or soften the immediate `price < 200-SMA` exit — it directly opposes the `close > SMA` entry and causes whipsaw re-entry every time price straddles the SMA; gate the regime exit on a confirmed close below the SMA over N bars, or drop it and rely on the EMA-stack break as the primary trend-exit. (2) Widen the trailing stop well above 8% (e.g. 15-25%) so the trailing stop stops truncating multi-week trends. (3) Add a re-entry cooldown / hysteresis (e.g. require a fresh EMA-stack cross, or wait K bars after an exit) so the strategy cannot re-enter on the very next bar. (4) Verify holding period lengthens to weeks and trade count drops toward the intended ~28-67 before re-submitting. Do NOT optimize the current version — parameter tuning cannot fix the structural churn, and the -68.8% churning result would only overfit.

Outcome Summary

Inheriting a promoted architecture does not inherit promotability — a long-only trend-continuation template only works on instruments that actually trend up, and porting it to a structurally bearish asset both loses money and degrades into whipsaw trading that violates the strategy's own multi-week holding premise.

The analyst abandoned it at the pre-optimization BACKTEST_REVIEW gate before any optimization ran, citing a catastrophic no-edge result plus a mechanism mismatch: the entry/exit stack whipsawed on consecutive bars rather than riding trends, and the root cause was a premise problem — a long-only trend-follower on a structurally bearish asset (FIL fell ~98% from its 2021 high) with no uptrend edge to tune toward.

Port the proven LINK/ATOM daily multi-week trend-continuation template to FILUSDT.BINANCE perpetual futures — a long-only, OHLCV-only strategy that enters when price is above its 200-day SMA regime filter and the 30/80 EMA stack is aligned up, aiming for 5-12 trend-riding entries per year for narrative diversification into decentralized storage.

The base backtest was a decisive money-loser: total return -68.8%, Sharpe -0.519, profit factor 0.649, negative expectancy of -$430/trade, and a 79.0% max drawdown (CI to 97.7%) across 175 long trades with a 45.7% win rate. It also failed to implement its own premise — producing ~30 entries/year with a ~1-day average hold instead of the promised 5-12 multi-week holds.

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