EthCoinMDailyTimeSeriesMomentumLS

ETH COIN-M Inverse Perp Medium-Horizon Time-Series Momentum, Long-Short (BINANCE_CM, OHLCV-Only, 2-Parameter)

A net-directional, single-instrument, medium-horizon TIME-SERIES MOMENTUM strategy on ETHUSD_PERP.BINANCE_CM (the ETH-margined inverse perpetual), using DAILY bars and pure OHLCV. The strategy holds ETH fully LONG when its trailing ~50-day return is positive (above a deadband) and fully SHORT when negative, flipping only on a confirmed trend-sign change. It deliberately leans into the ONE edge this session has repeatedly confirmed survives the recent tight-correlation regime: NET-DIRECTIONAL, MEDIUM-HORIZON (~50-60 day) momentum on the majors (the analyst's explicit finding that the directional tilt — not vol-scaling, not dollar-neutral relative momentum, not long-horizon 200d trend — is what stayed positive in 2024-2026). It fills the MOST under-represented, engine-CLEAN venue bucket: BINANCE_CM at 2.0% vs the >=5% target (the COIN-M Donchian siblings already validated to backtest_review, so there is no data-wall or sizing artifact here — unlike spot's CASH over-fill bug or Deribit's options marking bug). The inverse contract makes returns BTC/ETH-denominated with inverse convexity — a genuinely distinct return stream for the portfolio, not a re-expression of the existing USD-M BTC/ETH momentum signals. It also adds to the under-represented long_short bucket (14.5%). Only TWO core parameters (momentum lookback, deadband) to resist the overfitting that has recurred this session; no fade/reversion (anti-predictive), no options (engine-blocked), no vol-management overlay (falsified), no spot (CASH bug), no multi-leg basket (exposure-inflation / leg-feed bugs).

Implements the hypothesis exactly: single-instrument daily TSMOM on ETHUSD_PERP.BINANCE_CM, long-short, pure OHLCV, and only TWO searched parameters (lookback, deadband) - the sizing weight is underscore-prefixed so the optimizer and sensitivity engine skip it and the multiple-testing bar stays where the hypothesis wants it. calculate_signal returns the trailing return in natural units every bar (continuous, never gated behind a constant); should_enter/should_exit threshold it against the deadband. State is a single bounded deque with O(1) work per bar, so the ~2,100-bar daily history is trivially inside the sandbox budget. COIN-M correctness (L14/L66): position_size goes through contracts_for_notional, so the order is an integer contract count of fixed $10 USD face rather than coin units divided by price; leverage stays 1.0 and is read in sizing, so nothing inert is declared. On risk (L15) I made one deliberate deviation from the hypothesis's phrase 'fully long/short': the notional is 35% of equity, not 100%. My pre-submission dry run on the real ETHUSD_PERP daily catalog (2,118 bars, 2020-08..2026-06) shows this mechanism at 50% weight has a 52.6% max drawdown - at full weight that is ~100%, the exact compounding-blowup signature that has killed prior directional strategies here; at 35% the return stream is identical in shape (Sharpe is scale-invariant) with drawdown near 37%. Same dry run at the shipped defaults: 49 trades over 5.8 years (8.4/yr, avg hold 41 days), mean +9.4% per trade against a ~0.06% COIN-M round trip, so fees are a rounding error rather than the binding constraint. Importantly the edge is not one lucky cell - it is positive across the entire lookback x deadband grid I scanned (30/50/60/90 days x 0/2/5% deadband, per-trade means +3.8% to +18.2%), and the yearly path is positive in 2020, 2021, 2023, 2025 and 2026 with only 2022 (-4.6%) and 2024 (-3.3%) negative, which is the medium-horizon directional persistence the hypothesis is betting on. Honest caveat: base Sharpe on this family is ~0.6, so its ceiling is modest and it will live or die on whether that survives deflation - but the parameter surface is flat enough that the optimizer should not be able to manufacture a fragile peak.

drawdown_breach_single_day_artifact_ohlcv_tsm: ETH COIN-M daily TSMOM fails two independent disqualifiers at review. (1) L19 hard drawdown breach: max_drawdown 54.3% (>50%, CI high 87.9%) — a net-directional daily TSM that 2-param tuning cannot repair. (2) No measurable edge: only 48 trades in 5.8 years (L10 <100 tripwire) with Sharpe 0.426 and CI [-0.262, 1.077] straddling zero, and the entire result rests on ONE day — 2021-03-25 +65.1% (largest_win $65k) IS essentially all of 2021 (+73.7%) and most of the +112.7% total (kurtosis 10.4); strip it and 2022-2024 are flat-to-negative. It is closet-long ETH (benchmark_correlation 0.923, beta 0.52, information_ratio -0.632, alpha -0.039) — captured beta, not skill. avg_trade_return_pct 7.28% clears the CM fee floor so fees are not the constraint; the drawdown and unmeasurability are. This is the single-instrument OHLCV directional TSM family (0 survivors; the ETH-4H and BTC-daily vol-targeted TSM siblings just died overfit/no_edge in this session). A 225-trial / 2-param sweep on 48 trades whose PnL lives in one 2021 outlier would curve-fit the spike and die at DSR/holdout. Not worth the 2-hour optimization run.

Net-directional time-series momentum on the ETH COIN-M INVERSE perpetual (ETHUSD_PERP.BINANCE_CM, ETH-margined), DAILY bars, pure OHLCV, long-short. Each day it computes the trailing 50-day return of the close. Above +2% it holds a full LONG, below -2% a full SHORT; readings inside the deadband change nothing. The position is closed only when momentum flips THROUGH the opposite deadband, and the next bar re-establishes it on the other side - a confirmed stop-and-reverse whose hysteresis keeps a 50-day trend system from churning around zero. No stop-loss, no vol overlay, no supplementary data: the exposure is the trend sign and the size is a fixed 35% of equity notional, converted to an integer count of $10-face COIN-M contracts so the inverse contract is sized correctly and returns land in ETH.

At backtest_review weight DSR/PBO and walk-forward heavily; abandon before the full run if the flat grid doesn't survive deflation.

DEFLATION-SURVIVAL RISK (backtest-review analyst's call, not a code/fee block): single-instrument net-directional pure-OHLCV TSMOM — a deflation-fragile family — with base Sharpe 0.426, CI [-0.226, 1.111] straddling zero. For a directional trend follower base Sharpe is effectively the ceiling, so the 0.5 OOS promotion floor isn't comfortably cleared; the developer honestly notes the ~0.6 ceiling. Only 48 trades / 5.8yr against 2 searched params. NOT a no_edge death: the sandbox shows a real, fee-clearing, positive-expectancy edge (PF 2.56, +112%, avg_trade 8.16%, positive across the full grid and 5 of 7 years).

Risk officer should size for the >50% drawdown and inverse-contract convexity.

RISK: max_drawdown 54.3% (CI to 88.2%), end_unrealized -24.4% (large open loser at sample end). The 35% weight was chosen to avoid the ~100% blowup path but realized DD exceeds the 37% dry-run estimate. Disclosed characteristic of a full-directional momentum book, not a defect.

Treat as a directional momentum sleeve, not diversifying alpha.

alpha -0.039, IR -0.63 vs buy_hold with correlation 0.92 / beta 0.52 — correlated directional ETH exposure that doesn't beat buy-hold risk-adjusted. Unlike the vol-managed siblings this does NOT refute the hypothesis, which explicitly bets on net-directional momentum, not market-neutral alpha.

avg_trade_return_pct 7.28% clears the CM fee floor by a wide margin — not a fee death

Minimal 2-parameter design; correct inverse-contract sizing; balanced long/short (23/25)

profit_factor 2.4 nominally positive

L19 hard breach: max_drawdown 54.3% (>50%, CI high 87.9%) — structurally disqualifying, unfixable by 2-param tuning

Only 48 trades over 5.8 years (L10 <100 tripwire) — Sharpe unmeasured

Single-day dominance: 2021-03-25 = +65.1% (largest_win $65k) IS essentially the entire 2021 (+73.7%) and most of the +112.7% total; kurtosis 10.4

No measurable edge: Sharpe 0.426 with CI [-0.262, 1.077] straddling zero; flat-to-negative 2022-2024 once the spike is removed

Closet-long ETH: benchmark_correlation 0.923, beta 0.52, information_ratio -0.632, alpha -0.039 — captured beta, not skill

Single-instrument OHLCV directional TSM family (0 survivors; ETH-4H and BTC-daily vol-targeted TSM siblings died overfit/no_edge this session)

EthCoinMDailyTimeSeriesMomentumLS leaned into the session's one repeatedly-confirmed edge — medium-horizon net-directional momentum on the majors — implemented as a clean 2-parameter daily TSM on the under-represented, engine-clean ETH COIN-M inverse perp. The headline looked strong (+112.7%, PF 2.4, avg trade +7.28%), but the backtest review found it fails two independent bars: a 54.3% max drawdown breaching the L19 50% limit that 2-param tuning can't repair, and no measurable edge — 48 trades, a Sharpe CI straddling zero, and virtually the entire result resting on one +65% day in March 2021 with 2022-2024 flat-to-negative. With a 0.923 ETH correlation and negative information ratio it was captured beta, not skill. As another zero-survivor OHLCV directional TSM, it was abandoned at backtest review before any optimization or analysis.

A minimal 2-parameter net-directional daily TSM cannot escape the structural drawdown of a fully-long-or-short single asset (>50% breach) nor manufacture alpha — with ~8 trades/year the result is dominated by single outlier days and is captured beta (correlation 0.92, negative information ratio), not a distinct edge, however clean the venue.

It was abandoned at the pre-optimization BACKTEST_REVIEW gate on two independent disqualifiers: an L19 hard drawdown breach (54.3% > 50%, unfixable by 2-param tuning) and no measurable edge (only 48 trades, Sharpe CI straddling zero, PnL living in one 2021 outlier, and captured beta not skill). As another member of the single-instrument OHLCV directional TSM family (0 survivors), a 225-trial sweep on 48 trades would curve-fit the spike and die at DSR/holdout, so it was judged not worth the optimization run. Optimization, analysis, and risk review were never reached.

A net-directional, single-instrument, medium-horizon time-series momentum strategy on the ETH COIN-M inverse perp (ETHUSD_PERP.BINANCE_CM, daily bars, pure OHLCV) that held fully long when the trailing ~50-day return was above a deadband and fully short when below it, flipping only on a confirmed trend-sign change, with just two tunable parameters (lookback, deadband).

Over 5.8 years it made only 48 trades (23 long / 25 short) with total return +112.7%, profit factor 2.4, and avg_trade_return_pct 7.28% (clearing the CM fee floor), but no measurable risk-adjusted edge: Sharpe 0.426 with CI [-0.262, 1.077] straddling zero, and a 54.3% max drawdown (CI high 87.9%). The result was dominated by a single day (2021-03-25, +65.1% / $65k largest win) that is essentially all of 2021 and most of the total, with 2022-2024 flat-to-negative once stripped, and it was effectively closet-long ETH (correlation 0.923, beta 0.52, information ratio -0.632, alpha -0.039).