Strategies

Crowd-Positioning Timing, Crypto Restatement of Industrial Securities' Fund-Position Timing (BigQuant 2020-07-21): Fade Extremes in Binance Global Account Long/Short Ratio on 10 USD-M Alt Perps, Daily, 5-Day Clock, Long-Short

Outcome: Abandoned

BinanceAltCrowdPositioningFadeDaily

Outcome Summary

BinanceAltCrowdPositioningFadeDaily tried to adapt a Chinese mutual-fund positioning timing signal to crypto by fading crowd long/short ratio extremes on 10 Binance alt perps, based on a pre-study showing a weak gross edge of about +0.55% per trade. The code never passed verification. Layer 1.5 flagged a hypothesis/config mismatch over the optimization plan fixing leverage at 1.0, and the developer then resubmitted byte-identical code in iteration 3. The run was abandoned in a verification loop, so the backtest, optimization, analyst and risk stages were never reached.

Strategy report

Hypothesis

Classification: (a), a concrete mechanism on a market we cannot trade. The source measures the estimated equity allocation of Chinese public mutual funds and uses changes in that aggregate positioning to time the A-share index. It also splits time into trend and range regimes with an MA system and reports that the positioning signal only times the market in ranges. We have no fund-position data and no CNEQ estate to run it on (L103). The crypto analogue of 'aggregate positioning of an identifiable participant class' is Binance's published positioning metrics. The metrics feed (data/supplementary/metrics/<SYM>/all.parquet, BinanceVisionCollector, 5-minute rows from 2021-12-01) carries long_short_ratio (all-account global ratio) and top_trader_long_short_ratio. POINT-IN-TIME PRE-STUDY (daily closes, 10 alts, 2022-04 to 2026-09, non-overlapping 5-day events, |z|>1.5 on the 60-day z-score of log long_short_ratio): (1) FOLLOWING the crowd lost in every variant, and FADING it earned a gross +0.55% per 5-day trade over 1155 events (t ~1.9). By regime: range +0.44% (n=410, t 1.0), trend +0.62% (n=745, t 1.55). (2) The source's regime split is NOT reproduced: the signal is stronger in trends than in ranges. The hypothesis therefore uses no regime gate and records the source's range-only claim as falsified here. (3) Top-trader ratio and positioning-change (residualised on price) variants showed no edge, roughly 0 +/- 0.3%, and are excluded. (4) Fading crowd LONG extremes carries most of the edge (+1.02% on n=174 in ranges), while fading crowd SHORT extremes is about flat. (5) Per year, range-regime means are 2022 +0.45%, 2023 +0.92%, 2024 +0.33%, 2025 -0.02%, 2026 +0.62%. Seven of 10 symbols are positive. This is a modest, honestly weak edge. Implementation: one independent position per symbol, at most 10 concurrent, 10% of equity notional each, leverage fixed at 1.0. The universe is capped at 10 names per L173, and the per-bar work is O(1) rolling statistics on daily bars (precompute the daily metric series once in __init__ and use bisect for as-of lookup, per the sandbox-timeout lessons).

The source's fund-position factor maps to Binance's account long/short ratio, the only aggregate-positioning series we hold with multi-year history. The pre-study rejected the source's regime split and the top-trader and positioning-change variants, and kept the one variant with gross >4x cost. The trade is long-short (a 15.7% bucket against the <=55% long-only target). It uses 10 alts with little BTC/ETH corpus use and stays on BINANCE USD-M, not HL or COIN-M, per L168 and L172 (history length, pure-price mechanism). There is no calendar trigger (L179) and no inventory or grid (L171). The universe is capped at 10 names (L173). The 5-day clock comes from the pre-study's 5-day horizon, not from a time-in-market target (L166).

verification_loop: code_identical: iteration 3 resubmitted byte-identical code to iteration 2 — the previous feedback was not acted on, so the result cannot change. Either change the code to address it, or state recommendation='abandon' with the reason if no code change can help. Previous feedback: Verification failed (Layer 1.5 — hypothesis/config consistency) [class=hypothesis_mismatch]: - optimization_plan fixes leverage at 1.0

Implementation

Fades extreme Binance global account long/short positioning across ten altcoin perpetuals using daily signals and independent five-day holds.

Preserves the prior implementation and sets leverage to the research plan’s fixed 1.0; each leg uses 10% of equity notional.

Verification Results

code_identical: iteration 3 resubmitted byte-identical code to iteration 2 — the previous feedback was not acted on, so the result cannot change. Either change the code to address it, or state recommendation='abandon' with the reason if no code change can help. Previous feedback: Verification failed (Layer 1.5 — hypothesis/config consistency) [class=hypothesis_mismatch]: - optimization_plan fixes leverage at 1.0

Iteration History

Verification failed (Layer 1.5 — hypothesis/config consistency) [class=hypothesis_mismatch]: - optimization_plan fixes leverage at 1.0

Outcome Summary

When a hypothesis/config consistency check fails, change the code or config to fix the named mismatch (here, how leverage is fixed relative to the optimization_plan), or explicitly recommend abandonment; resubmitting unchanged code ends the run.

It was abandoned at verification after 3 iterations. Layer 1.5 (hypothesis/config consistency) flagged a hypothesis_mismatch ('optimization_plan fixes leverage at 1.0'), and iteration 3 resubmitted code byte-identical to iteration 2, so the verification loop stopped it.

A crypto version of a fund-positioning timing study: fade extremes (|z|>1.5 on the 60-day z-score of log long_short_ratio) in Binance's all-account long/short ratio across 10 USD-M alt perps on daily bars, holding each position for 5 days, long or short, at 10% of equity per symbol and 1.0 leverage.

No backtest or optimization was produced. The only numbers come from the hypothesis's own pre-study: a gross +0.55% per 5-day fade trade over 1155 events (t ~1.9), described there as a modest, honestly weak edge.

Analysis

利用基金仓位信息对市场择时-兴业证券-20200721 - BigQuant量化交易

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