BtcDvolRegimeTrendContinuationLS

BTC Implied-Vol-Regime Trend Continuation, Long-Short (Trade the BTCUSDT.BINANCE Perp WITH the Price Trend, Gated by the DERIBIT Implied-Vol (DVOL) REGIME — NEW Information Source: Ride Calm/Falling-IV Uptrends & Stressed/Rising-IV Downtrends, ATR-Trailing, 3-Parameter)

A LONG-SHORT directional strategy that TRADES the liquid BTCUSDT.BINANCE USD-M perpetual WITH the price trend, but gates and biases that trend by a genuinely NEW INFORMATION SOURCE — the DERIBIT implied-volatility (DVOL) REGIME — per the meta-learner's L38 mandate. It is deliberately aligned with the one edge that WORKS in this factory (momentum/trend — the paper-stage survivor) rather than another contrarian FADE, because every BTC fade this session died (liquidation, OI-divergence, false-breakout, trend-aligned-MR, deleveraging), and my three pending options ideas are all extreme-fades likely to share that fate. The distinct thesis: FORWARD-LOOKING implied vol tells you when a trend is safe to ride versus fragile. Calm, LOW/FALLING implied vol accompanies healthy, persistent uptrends (leverage building smoothly) → ride LONG; STRESSED, HIGH/RISING implied vol accompanies disorderly downtrends and vol-expansion → ride SHORT; and it stands aside when the price trend and the IV regime disagree (e.g. price up but IV spiking = fragile rally). This differs from the DEAD realized-vol-regime-gated momentum (that used backward-looking realized vol; this uses forward-looking OPTIONS-implied vol, a new information source) and from my pending options-extreme FADES (this is trend CONTINUATION, not a fade). It is NOT an options strategy (no option legs → dev-safe; DVOL is a signal input only). It stays on BTC (deepest options market, the one tradable instrument), single perp execution, pure trend + a regime gate → measurable (~100-200 trades), fee-clearing (BTC trend legs ~2-5% vs ~0.10% fee), and fills the under-target LONG-SHORT bucket (13.5% vs 86.5% long-only). Options data has adequate multi-year history (unlike L/S-ratio/taker coverage ceilings). 3 parameters.

The hypothesis's key judgement is that this should ride trends rather than fade extremes, and I agree with the evidence behind it: every BTC fade this session — liquidation cascade, OI divergence, false breakout, trend-aligned reversion — died the same way, with the extreme extending, while trend is the one family with a survivor. So the new information source is used as a GATE on a continuation trade, and the gate is genuinely directional rather than symmetric: an uptrend needs calm or falling implied vol, a downtrend needs stressed or rising implied vol, and price advancing into a vol spike is explicitly excluded as a fragile rally. That asymmetry is also what separates this from the dead realized-vol-regime version, which used a backward-looking and direction-free vol measure. Three design decisions carry the risk, and each is a lesson paid for elsewhere in this session. First, the gate is a GRADED weight, not a second threshold: trend x clip(|stress|/ref) lets the two dimensions trade off continuously, because ANDing two independent tails has repeatedly collapsed sibling trade counts to single digits and the multiplicative version of the same mistake (a product of several 0-1 factors) made another sibling literally untradeable — I built the graded form in from the start rather than after a zero-trade iteration. Second, the DVOL parser is hardened up front: keys are coerced whether they are ints, floats or numeric strings and samples are keyed by the resulting nanosecond value, which handles int-only, string-only and dual-key shapes without double-counting. Skipping string keys is correct for funding_rates but silently discarded an entire string-keyed DVOL series on a sibling and cost two zero-trade iterations, so the parse log reports rows, span and IV range and on_stop reports bars_with_iv plus the observed min/max IV rank — the three numbers that distinguish 'no data' from 'gate too tight'. Third, risk: the trailing stop is detected INTRABAR and booked at the bar close (close-only stops turned modeled risk into 5-10x overruns on two siblings), sizing is gap-aware so a tight trail cannot silently buy a maximum position, and losing the IV gate deliberately does NOT flatten a live position — the gate decides when risk is taken, the trail decides when it is given back, and both non-aligned branches clip strictly below the exit level so a regime flicker cannot fire a hair-trigger exit. One honest caveat for the Analyst: the biggest open question is coverage, not logic. If deribit_dvol turns out to have thin or absent history in this pipeline, this strategy will correctly trade nothing rather than fall back to price-only trend, and the diagnostics will say so plainly — that would be a data-availability verdict on the whole options-informed family rather than a fixable defect in this strategy.

Decisive overfit with a NEGATIVE out-of-sample edge: four UNWAIVABLE hard gates — walk-forward is_overfitted=TRUE (IS 7.51 → OOS -2.12), OOS Sharpe -2.12 ≤ 0, PBO 0.6918 > 0.5, holdout FAILED (Sharpe -0.30, ratio 0.0). Deflated Sharpe 0.000 with IS trial Sharpes up to 12 (expected-max 5.43) = pure best-of-N noise; sensitivity FAILED (2 cliffs), sample collapsed to 43 trades (kurtosis 94). The OOS doesn't just decay — it inverts negative (-2.12 WF, -0.30 holdout). The Deribit-DVOL/implied-vol source is exhausted: 6th DVOL/IV directional strategy to die across fade and trend-gate constructions. No robust region to tune.

Long/short trend continuation on the BTCUSDT.BINANCE USD-M perpetual, 4-HOUR bars, gated by the Deribit DVOL implied-vol regime; no option legs. The trend leg is the close-to-close change over mom_lookback bars in random-walk sigma units (change / (ATR x sqrt(n))). The regime leg is a signed stress score in [-1,1] combining where DVOL sits in its own trailing 540-bar percentile distribution with its relative change over the last 6 bars — negative means calm and/or falling, positive means stressed and/or rising. An uptrend is only ridden when stress is negative and a downtrend only when stress is positive; the signal is trend x clip(|stress|/stress_ref, 0, 1), so a more decisive regime lets a smaller trend qualify and vice versa, and on disagreement (price up into a vol spike, price down with vol collapsing) the signal is damped and clipped strictly below the exit level so it can neither open nor close a position. Entry is a fresh cross of +/- entry_thresh. Exits are an ATR trailing stop trail_atr_mult ATRs from the best close since entry, detected intrabar and booked at the bar close, and the signal reaching -exit_frac x entry_thresh against the position; there is no fixed target. Sizing risks a locked 2% of equity against max(trail_atr_mult, gap_atr_mult) x ATR, capped at 0.5x equity notional. If DVOL is missing or stale no position is taken. Three tunables: mom_lookback, entry_thresh, trail_atr_mult.

At BACKTEST_REVIEW / walk-forward: (1) compare the DVOL-gated trend against the UNGATED trend over the same span -- if the gate does not lift risk-adjusted return or cut drawdown, the options-IV source is not earning its place and this collapses to the plain trend survivor; (2) verify the edge is not one-trend-dependent (avg_trade stays above the fee floor with the top 1-2 wins removed, tail recurs across windows); (3) confirm OOS Sharpe > 0, IS/OOS ratio < 3.0, holdout >= 70% of WF-OOS Sharpe.

DOES THE DVOL GATE ADD VALUE? -- the No.1 item for the analyst, because it is what justifies the new information source. This clears every hard gate (avg_trade_return_pct 1.049% at 7x the fee floor, PF 1.52, capacity_usd $70.3M, DD 7.3%, positive skew 0.85), but the trend leg alone is the factory's known survivor, so the open question is whether the DVOL regime gate IMPROVES the trend or merely thins it. The developer's aligned_long/aligned_short/disagree and above-threshold diagnostics enable the direct test. Secondary robustness flags at n=22: Sharpe 0.515 has a CI [-1.20, 2.01] straddling zero, and the +7.09% is somewhat tail-weighted (largest_win $4,280 ~= 60% of the $7,093 net profit) -- though, unlike the failed one-trade-dependent cases, removing the single largest win leaves the year net-POSITIVE, and positive skew with a favorable RR (avg_win $2,302 vs avg_loss $1,048) is the EXPECTED signature of a trailing-stop trend follower, not a defect.

Confirm the multi-year trade count from the on_stop diagnostics (aligned/above-threshold/entry counts) is at or above ~100 with both long and short legs firing (long 9 / short 13 here is balanced). Optionally align the IV-rank computation with the sibling's series-based approach for cleanliness, though it is not blocking. Read iv_rank_min/max to confirm the gate spans the distribution.

BORDERLINE MEASURABILITY + a minor plumbing note. 22 trades over 363 days projects to ~130 over a multi-year BTC 4H span -- above the ~100 floor but not a large sample, and short of the developer's own '25-40/year' projection. Separately, note the IV rank here is computed from a BAR-SAMPLED _iv_hist (gated at 100 samples), which the DVOL-fade sibling switched AWAY from (to a series-based rank) to remove bar-warmup coupling; here it works (22 trades, not a zero-trade failure) and only costs a ~17-day warmup, but it is slightly less clean than the series-based version and its window depends on bar cadence. impact_cost_pct 3.77% / capacity $70.3M are healthy (contrast the DVOL-fade's $64K collapse).

Positive per-trade edge above fees: avg_trade_return_pct +0.91% (~9x round-trip cost), profit_factor 1.36, expectancy +$310/trade, favorable RR (~2.04x)

Genuinely new forward-looking signal (Deribit DVOL regime) per the L38 mandate — data parses and fires, a distinct hypothesis not a price re-parameterization; trend-continuation (the surviving family), two-sided (54 long / 46 short), low max_drawdown 14.3%

Measurable sample at the floor (100 trades) on BTC, the proven instrument

Sharpe 0.376 with CI [-0.278, 0.997] straddling zero — weak; likely to face robustness (PBO/holdout/DSR) headwinds that killed similar-Sharpe BTC trend-gate siblings

Asymmetric edge: short_win_rate 0.26 vs long_win_rate 0.52 — the stressed-IV-downtrend leg barely works, most edge is the long side; 2022 was -11%

100 trades is right at the measurability floor; kurtosis 25.8 indicates some tail dependence

BtcDvolRegimeTrendContinuationLS deliberately aligned with the session's one surviving family — trend continuation — while adding a genuinely new forward-looking signal, using the Deribit DVOL implied-vol regime to decide when a BTC trend was safe to ride versus fragile. Its initial backtest cleared fees (100 trades, +32.5%, profit factor 1.36, avg trade +0.91%) but was weak and asymmetric, with a Sharpe of 0.376 whose confidence interval straddled zero and a short leg that barely worked. Optimization was decisive and brutal: the out-of-sample edge didn't merely decay but inverted negative (walk-forward OOS -2.12, holdout -0.30), PBO hit 0.69, and the deflated Sharpe was zero against IS trial Sharpes as high as 12 — textbook best-of-N noise. The analyst abandoned it, noting this was the sixth DVOL/implied-vol directional strategy to die and that the information source now looks exhausted.

A marginal in-sample Sharpe (~0.38) whose CI straddles zero and whose edge is one-sided (the short/stressed-IV leg barely worked) is a poor optimization candidate — and the Deribit-DVOL/implied-vol source is now exhausted, having died across six fade and trend-gate constructions this session.

Backtest-review passed it to optimization ('optimize'), but the analyst abandoned it there on four unwaivable hard gates: walk-forward flagged overfit (OOS Sharpe -2.12 ≤ 0), PBO 0.6918 > 0.5, holdout failed (Sharpe -0.30, ratio 0.0), and deflated Sharpe 0.000 with IS trial Sharpes up to 12 marking it as pure best-of-N noise; sensitivity also failed with two cliffs.

Ride the BTCUSDT.BINANCE perp WITH the price trend but gate it by the Deribit DVOL implied-vol regime — going long in calm/falling-IV uptrends, short in stressed/rising-IV downtrends, and standing aside when price trend and IV regime disagree.

The initial backtest was marginal but fee-clearing: 100 trades (54 long / 46 short), total return +32.5%, Sharpe 0.376 (CI [-0.278, 0.997] straddling zero), profit factor 1.36, avg per-trade return +0.91% (~9x fees), max drawdown 14.3% — with a lopsided edge (long win rate 52% vs short win rate 26%). Optimization then inverted: walk-forward IS Sharpe 7.51 collapsed to OOS -2.12 and holdout Sharpe -0.30, with the sample shrinking to 43 trades.