OiInflowWeeklyRotationMajorsLongSpot

Open-Interest-Inflow Long-Only Spot Rotation over Majors (Weekly Rebalance, Signal from Perp OI Growth + Uptrend Regime Gate, Execution on BINANCE_SPOT, 3-Parameter)

A LONG-ONLY, multi-instrument WEEKLY rotation strategy executed on Binance SPOT majors, whose ranking factor is a genuine non-price futures feed with confirmed multi-year history: perpetual open-interest growth. Each week, rank a fixed basket of liquid majors by the trailing growth in their perp open interest (fresh capital / conviction inflow), and hold the top-3 that are ALSO above their long-term moving average (regime gate to avoid buying inflows into a crash). This is deliberately in the PROMOTED-WINNER family (long-only rotation) rather than the 0/53-survival dollar-neutral basket class (L12) — turnover is low, only one long leg per name, no shorting, so per-rebalance fees are a fraction of the market-neutral rank-and-rotate that repeatedly died. Critically, the SORT factor is OI inflow, NOT relative strength — the exact 'new signal input no promoted basket already uses' that L24 requires to avoid the duplicate-momentum-rotation graveyard. It is a completely different family from the abandoned taker-flow divergence reversal it replaces: multi-instrument not single-name, weekly not intraday, long-only spot not long-short futures, rotation not fade, OI-inflow factor not order-flow-vs-price divergence. It fills the highest-survival, most under-represented venue (BINANCE_SPOT at 6%) while sourcing an edge from a futures-only microstructure feed — a real cross-venue signal/execution split.

The QA block was that the hypothesis's required factor — per-name perp OI growth — was absent, replaced by dollar-turnover growth, on the belief that per-instrument OI cannot be injected. That belief is wrong, and the previous file even hints at it: it already listed "open_interest_by_instrument" in _supp_keys but only ever read supp["open_interest"] (the primary series). The per-leg map is built by subprocess_entry._add_per_leg_open_interest for every leg whenever a strategy references the key, and supp_collector._OI_VENUES includes BINANCE_SPOT, which maps each spot leg to its matching USDT perp. I verified the data for all ten basket members: BTC 28,923 hourly points from 2020-09, the other nine 37k-41k from 2021-12, every one current to 2026-09. So the ranking factor is now the real thing: per-name OI inflow, with the close*volume turnover proxy deleted outright (the only remaining mentions of volume are docstring prose explaining the removal). This also repairs QA item 3: because every leg carries OI, the market regime gate is now the growth of basket-aggregate perp OI rather than degrading to BTC-above-its-MA for most of the window, and when OI history is insufficient the book stays FLAT rather than falling back to a price proxy. The per-name long-term MA gate stays because the hypothesis itself specifies it ('hold the top-3 that are ALSO above their long-term moving average'). I checked the factor actually differentiates before submitting: replayed over the real per-name OI series it gives 245 weekly rebalances, an average of 5.55 of 10 names with OI building, and 1.18 name changes per rebalance (~290 sleeve trades) — a live, low-turnover rank, not the alphabetical-tiebreak collapse a single shared series produced. I also fixed the uptime-relative cadence ((self._proc - 1) % rebalance_period) to a calendar bucket derived from the bar timestamp, per the schedule rule, so rotation dates survive restarts and replay. Point-in-time discipline is unchanged (np.searchsorted as-of, no abs() matching), buffers are bounded, and _param_bounds now mirrors every clamp. On the advisory: the previous +1226% run was long-only majors beta, and with the rank now driven by OI rather than turnover the return profile should be judged fresh — I have deliberately not tuned anything toward that old number.

L45/L49/L19 long-only multi-instrument majors basket carried by crypto beta with NEGATIVE alpha over its own benchmark — not worth 2 hours of optimization. Despite the correctly-wired OI signal, information_ratio is -0.948 versus the equal-weight-basket benchmark (benchmark_meaningful=true): the OI sort subtracts risk-adjusted value over a naive equal-weight hold of the same 10 names, so the +19.8% headline is long-crypto beta (beta 0.12), not signal. It carries a 64.75% max drawdown (CI to 84%), above the 50% hard-abandon bar (L19), and the OI-growth regime gate failed to cut the 2022 crash (-56.8% year, luna -30.5%, ftx -28.9%, rate-shock -42.5%). It is not significant: Sharpe 0.19 with bootstrap CI [-0.456, 0.826] straddling zero, PSR 0.643, and over ~9 years it returned only +19.8% (CAGR 3.7%) — underperforming buy-and-hold with stressed-regime Sharpe -0.36. The base Sharpe is the ceiling the ~225-trial deflation only lowers; there is no parameter region that turns a beta stream with negative alpha and a 65% drawdown into an edge. Failure pattern: no_edge/risk_reject long-only majors rotation basket, beta-carried with negative IR vs benchmark, >50% drawdown (L45/L49/L19).

Long-only weekly cross-sectional rotation over 10 liquid Binance SPOT majors (BTC, ETH, SOL, BNB, XRP, ADA, DOGE, LINK, LTC, AVAX) on daily bars, ranked by each name's own perpetual OPEN-INTEREST growth — the log change between the mean of the last `lookback` daily OI samples and the mean of the prior `lookback` samples, i.e. fresh leveraged capital flowing into that name. Signal is sourced from the USD-M perp feed while execution is on spot, the cross-venue split the hypothesis specifies. Each week the book holds the top_n names whose OI is building AND that trade above their own long-term moving average (crash filter), equal-weight at 31% of equity each. A market-wide gate on basket-AGGREGATE perp OI growth keeps the book flat whenever fresh money is leaving the complex. Rebalances fire on a calendar-anchored weekly grid derived from the bar timestamp; names held across consecutive weeks are never re-traded, so only the delta pays spot fees. Leverage 1.0, no shorts.

Verification failed (Layer 4 — QA review) [class=hypothesis_mismatch]: - [edge_concern] The strong sandbox (+1226%, Sharpe 0.95, avg_trade_return_pct 9.47%, 232 trades) is long-only majors beta/trend, not evidence of the hypothesized OI edge: benchmark_correlation 0.53, beta 0.27, return concentrated in bull terciles, worst losses in 2022 rate-shock/FTX, end_unrealized -61.7%, max_drawdown 44%. It measures a turnover+trend rotation, not the OI-inflow factor. Advisory; the block is fidelity. - Core edge absent: the hypothesis ranks by per-name perp OPEN-INTEREST growth (the explicitly-required novel futures-only signal, L24). Per-instrument OI is never injected (primary only), so the developer substitutes the RANK with per-name dollar TURNOVER growth (close*volume) — an OHLCV participation measure, not OI. The specified OI-inflow ranking edge and its cross-venue signal/execution split are gone. - OI barely participates even in its demoted role: _market_risk_on degrades to BTC-above-MA (pure price) for most of the window per the developer's own disclosure that OI covers only a fraction of history. The executable strategy is a spot-majors volume+trend rotation, not the OI-inflow rotation described. Same per-name-metrics infeasibility that failed the OI-fragility / retail-crowding / top-trader-rotation siblings.

Size new entries from free cash, or exit dropped names in a first pass before submitting entries in a second pass.

New entries are sized at equity*per_name_frac(0.31)*leverage using TOTAL account equity, but on a CASH (spot) account most equity is already tied up in persisting positions. When rotating in a new name before the dropped name's exit fill settles, available cash may be < the computed notional and the BUY can be rejected. Not a signal-correctness bug, but can under-fill rotations.

Log the count of legs with non-empty OI series at startup so a degradation to primary-only is visible in diagnostics.

The mechanism depends on infrastructure injecting open_interest_by_instrument for every spot leg. If that per-leg map is absent on the real-backtest path, the primary-only fallback (supp['open_interest']) leaves only BTC rankable and the basket rotation silently collapses to a BTC-only book. The sandbox producing 196 trades indicates OI history was present (the gate requires it), but the full run should be confirmed to rank all 10 names.

The iter-2 fix genuinely wires the per-name perp OI feed (a real non-price futures signal), so this is not a duplicate-momentum-rotation and the OI ranking is implemented as the hypothesis intends.

avg_trade_return_pct 2.01% clears the spot fee floor, and commissions are only 3.66% of gross, so this is NOT a fee-edge failure.

max_drawdown 64.75% (CI to 84%) — above the 50% hard-abandon bar (L19); 2022 annual return -56.8%, luna window -30.5%, ftx -28.9%, rate-shock -42.5%. The OI regime gate did NOT protect against the crash.

Negative alpha and negative information_ratio (-0.948) versus its OWN equal-weight-basket benchmark (benchmark_meaningful=true): the OI sort adds NO risk-adjusted value over a naive equal-weight hold of the same 10 names — the +19.8% headline is just long-crypto beta (beta 0.12, corr 0.33).

No significance: Sharpe 0.19 with bootstrap CI [-0.456, 0.826] straddling zero, PSR 0.643. Over ~9 years (2017-2026) it made only +19.8% total (CAGR 3.7%), losing to buy-and-hold.

Return is concentrated in single bull-regime spikes (2024-11-21 +54.9% day, 2024 +90% year) while stressed-vol regime Sharpe is -0.36; recent regime is flat-to-negative (2025 +2.1%, 2026 +1.1%).

BtcPerpOiInflowTrendLongSpot

Verification failed (Layer 4 — QA review): - WHOLE-STRATEGY SWAP — the code does not implement THIS hypothesis. The hypothesis is a MULTI-INSTRUMENT, WEEKLY, CROSS-SECTIONAL rotation: 'rank a fixed basket of liquid majors by the trailing growth in their perp open interest ... and hold the top-3 that are ALSO above their long-term moving average.' The code is a SINGLE-NAME (BTCUSDT.BINANCE_SPOT only, no extra_instruments), 4-HOUR, long/flat timing overlay on one asset. Three defining elements of the hypothesis are all absent: (1) the basket / top-3 selection — there is no universe and no ranking across instruments; (2) the cross-sectional RANK factor — the code z-scores one name's own OI against its own history, which is a time-series timing gate, not the cross-sectional 'rank the basket' mechanic the hypothesis names; (3) the WEEKLY rebalance — config bar_type is 4-HOUR. Per the QA checklist, when a cross-sectional-basket hypothesis's code fails to rank N instruments, and instrument set + timeframe + mechanism all differ, that is a critical mislink/mis-generation. The developer's rationale openly concedes the substitution ('A literal multi-leg cross-sectional OI rotation ... cannot obtain OI for its extra legs ... I therefore implemented the edge as a single-name ... overlay'). - UNMEASURABLE trade count driven by an OI data-coverage limitation. The sandbox produced only 8 trades over 340 data_days, and the developer states this is 'capped at ~8 purely by the 56-day OI coverage.' Eight trades cannot distinguish edge from noise: Sharpe 0.668 with CI [-1.03, +2.18] straddles zero, return_kurtosis 54 and skew 3.27 indicate one or two trades dominate the result, and avg_loss ($2,488) actually exceeds avg_win ($1,878) despite the 75% win rate. Per prior QA learnings (a gate implying <~100 trades over the backtest span is unmeasurable), this must be rejected at Layer 4 rather than burning backtest/optimization iterations on a result that is structurally noise. It also contradicts the hypothesis's central claim of 'confirmed multi-year history / OI catalogued back to ~2021' — the actual on-disk OI is ~56 days, so the premise that this edge can be measured over a multi-year span does not hold with available data.

OiInflowWeeklyCrossSectionalRotationMajorsLongSpot

Verification failed (Layer 4 — QA review): - The cross-sectional OI-inflow rank — the strategy's entire reason for existing — is non-operative in the executed backtest. Only the PRIMARY (BTC) OI series is injected; `open_interest_by_instrument` is absent, so in __init__ every non-primary name falls back to BTC's OI series. `_oi_growth` then returns an IDENTICAL value for all 10 names, and `scored.sort(key=lambda x: (-g, symbol))` collapses to an ALPHABETICAL tiebreak. The executed strategy is 'when BTC OI is rising, hold the 3 alphabetically-first MA-passing majors' — an arbitrary basket, not OI-inflow rotation. The 'new signal input L24 requires' does nothing to differentiate names. - Decisively unmeasurable: 7 trades over 363 days, Sharpe CI [-1.61, +1.62] straddling zero. The config rationale itself admits OI coverage is only ~56 days, so the strategy is flat for ~85% of the window — the same data wall that killed the sibling OI/positioning families. - avg_trade_return_pct 0.0517% does not clear the ~0.20% spot round-trip / 0.25% floor; positive dollar result is a single fat tail.

OiRegimeTurnoverInflowWeeklyRotationMajorsLongSpot