LinkSpotRegimeHoldLong

LINK Spot Long-Hold with 200-SMA Regime Filter (Daily, Pure Regime-Flat Exit)

A long-only single-instrument PURE regime-conditional buy-and-hold strategy on LINKUSDT spot (BINANCE_SPOT) using daily bars and OHLCV-only data. This applies the SIMPLEST POSSIBLE mechanism in the pipeline (single 200-SMA regime filter, binary regime-flat exit, one tunable parameter) to LINK as the third asset for this proven architecture class (after EthSpotRegimeHoldLong and SolSpotRegimeHoldLong, both pending). Critically, this is NOT an architecture-mirror of a complex strategy that has failed across assets (e.g., 21/55 EMA, 50/200 Golden Cross, 30/80 Multi-Week Trend, BTC Spot Drawdown). The Pure Regime Hold mechanism has structurally different failure modes than complex daily-trend strategies: it cannot have outlier-trade dependency (it holds for FULL regime duration, capturing many days per cycle), it cannot have parameter cliffs (only 1 parameter), and it cannot have ATR-trail volatility-calibration issues (no trail). The portfolio impact is: LINK currently has only ONE working strategy (LinkDailyMultiWeekTrendContinuationLong perp Sharpe 2.55). This proposal adds a SPOT-venue + binary-regime-exit mechanism that is mechanism-orthogonal (state-based EMA continuation vs binary regime change) AND venue-orthogonal (perp vs spot). Asset selection is justified by LINK-SPECIFIC empirical evidence: LINK has demonstrated clean response to daily 200-SMA regime filtering via the Multi-Week Trend success. The simple mechanism extracts only the proven element (regime classification) and discards the complex EMA-stack machinery that other LINK daily strategies (Golden Cross, 21/55 EMA, 4H Volume Breakout) failed on. NO bait-and-switch — strictly LINKUSDT.BINANCE_SPOT throughout. ONE tunable parameter only (SMA period) means cliff_count cannot exceed 1.

Iteration 3 minimal fix for the Layer-3 backtest artifact. Root cause (per feedback): the previous should_exit returned True on a bad bar, routing a market SELL that filled at the corrupt spiked 2021-05-22 close and booked a phantom +375%/day gain that dominated the backtest. Fix: on a flagged bad bar the strategy performs NO transaction at that bar's price -- should_enter already returned None, and should_exit now returns False, carrying the prior position and prior clean price. Because calculate_signal never appends the corrupt close, _closes[-1] and _sma remain clean, so the regime exit only fires on a subsequent clean bar at a clean price. This is a single-branch change (return True -> return False plus docstring/comment updates); imports, param plumbing, and the continuous tanh(edge) signal are untouched, preserving the already-passing Layer 1 and Layer 2 checks. After this fix no single daily return should exceed a plausible long-only-50% bound and annualized volatility should fall to a sane range; if the clean re-run shows no genuine regime edge (negative alpha vs buy-hold, drawdown not materially below buy-hold), the analyst should abandon rather than optimize.

Not worth optimizing. (1) The +375% single-day artifact (2021-05-22, and a +46% day 2024-04-13) that the bad-bar guard was written to eliminate is STILL present in daily_returns — a +375% portfolio day at 50% sizing implies an impossible +750% asset day, so every headline metric (Sharpe 2.84, ci_low -22.9, vol 1359%) is contaminated; the developer already iterated to fix this exact artifact and it failed. (2) Independently, and decisively, the regime-hold produces NEGATIVE alpha (-0.041) and NEGATIVE information ratio (-0.694) versus a meaningful buy-hold benchmark even WITH the artifact inflating returns — removing the phantom gains makes it strictly worse — alongside a 77.5% max drawdown. The single tunable parameter (sma_period) cannot create alpha or fix a marking artifact, so 2 hours of optimization cannot rescue a mechanism that subtracts risk-adjusted value on LINK. Symptom stated, root cause of the artifact not independently verified — recommend the mechanism not be re-attempted on LINK without evidence that regime filtering beats buy-hold on a clean return series.

Long-only LINKUSDT spot buy-and-hold gated by a 200-SMA regime filter with a symmetric hysteresis dead-band. Enters when daily close clears SMA*(1+band), exits when close falls below SMA*(1-band). Holds for the full regime duration with no stops/TP/trail. A bad-bar guard now CARRIES the position through anomalous/spiked daily bars instead of selling into them, so no order is ever transacted at a corrupt bar price.

Simple, single-parameter mechanism with no leverage, no shorts, no liquidation risk

Sufficient trade sample structure in principle (20 full-regime holds over 7 years, exposure 52%)

Persistent +375% single-day artifact (2021-05-22) that the bad-bar guard was explicitly written to remove — plus other implausible days (2024-04-13 +46%) — makes all headline metrics untrustworthy (Sharpe 2.84 rests on these; sharpe_ci_low -22.9, annualized_vol 1359%)

Negative alpha (-0.041) and negative information ratio (-0.694) vs a meaningful buy-hold benchmark EVEN WITH the inflating artifact — the regime filter subtracts risk-adjusted value; de-artifacting makes it worse

77.5% max drawdown, 1622-day drawdown duration — no risk control benefit from the regime filter

Return concentrated in outlier wins (win_rate 15%, avg_win 128k vs avg_loss 16k) rather than the 'many days per cycle' the hypothesis claims

artifact-inflated (ci_low -22.9)

clean

The bad-bar guard did not eliminate the +375%/day artifact — 2021-05-22 still shows +375.62% and drives the whole backtest. Diagnose and fix before any optimization: (1) The guard's should_exit=True-on-bad-bar path almost certainly SELLS the open long at the corrupt spiked price, booking a phantom gain. On a bad bar you must NOT transact at that bar's price at all — either skip the bar entirely (carry the prior position and prior price, submit no order until a clean bar), or if you must flatten, do it on the NEXT clean bar at a clean price. (2) Confirm by reading the detail_file trade for the exit dated on/around 2021-05-22 — check the fill price vs the surrounding closes; a fill far above neighboring bars confirms the spike-sale. (3) Verify LINKUSDT.BINANCE_SPOT data on 2021-05-22 for a bad print; if the data itself is spiked, the fix is data-side filtering, not just a trade guard. After the fix, re-run and confirm no single daily return exceeds a plausible bound (~|30-40%| for a 50%-sized long-only spot book) and that annualized_volatility drops to a sane range. NOTE: even once the artifact is removed, the visible edge is negative-alpha vs buy-hold with a 77% drawdown — if the clean re-run does not show a genuine regime edge (positive alpha/IR, drawdown materially below buy-hold), abandon rather than optimize.

LinkSpotRegimeHoldLong applied the simplest possible mechanism — a single 200-SMA regime filter with a hysteresis band and bad-bar guard — to buy-and-hold LINK spot only during up-regimes, betting that a proven regime-classification edge would transfer cleanly to this asset. The backtest looked superficially strong (107.3% return, 13.8% CAGR, Sharpe 2.84 over 20 trades) but rested on a persistent +375% single-day artifact the developer had already iterated three times to fix, leaving vol at 1359% and the Sharpe confidence interval bottoming at -22.9. Even with those phantom gains inflating results, the strategy delivered negative alpha (-0.041), a negative information ratio (-0.694), and a 77.5% drawdown against a meaningful buy-hold benchmark. The analyst abandoned it at the pre-optimization backtest-review gate, concluding that one tunable parameter could neither create alpha nor fix the marking artifact, and it never reached optimization or promotion.

A single-parameter regime filter cannot manufacture alpha or repair a data-marking artifact, so a regime-hold mechanism should not be re-attempted on LINK without first proving it beats buy-hold on a clean, de-artifacted return series.

At the backtest-review gate the analyst returned an 'abandon' verdict: a +375% single-day marking artifact (2021-05-22) that the bad-bar guard was written to remove was still present, contaminating every headline metric, and independently the regime filter subtracted risk-adjusted value (negative alpha and information ratio) — so it was not worth the 3-phase optimization and never reached optimization or the post-optimization analyst stage.

A long-only LINKUSDT spot buy-and-hold strategy that stays invested only while price is above a 200-SMA regime filter (with a hysteresis band and a bad-bar guard) and goes flat on a confirmed regime flip.

Over a 2019-2026 backtest it produced 107.3% total return, 13.8% CAGR and a headline Sharpe of 2.84 across 20 trades, but with only a 15% win rate and a 77.5% max drawdown. Crucially the Sharpe was artifact-inflated (sharpe_ci_low -22.9, annualized vol 1359%) and it posted negative alpha (-0.041) and negative information ratio (-0.694) versus a buy-hold benchmark.