Short-Horizon Cross-Sectional REVERSAL as a Liquidity-Provision Premium on 12 High-Idiosyncratic-Vol USEQ Large Caps — Dollar-Neutral Long/Short, Vol-Scaled 2-Session Ranking, Non-Overlapping 2-Session Hold, News-Jump Exclusion, DAILY Bars (commission-free venue is the whole point)

A DOLLAR-NEUTRAL, LONG-SHORT, CROSS-SECTIONAL SHORT-HORIZON REVERSAL strategy on the commission-free USEQ venue, traded on 1-DAY bars across a FIXED, pre-declared 12-name universe of sector-diverse, high-idiosyncratic-volatility S&P 500 large caps: AMD, MU, WDC (semis/hardware), NFLX (media), BA, CAT, FDX (industrials), GS, SCHW (financials), HAL, DVN (energy), PFE (healthcare). Catalog verified: every name has a 1-DAY bar directory; the binding common start is NFLX (2002-05), so the usable joint window is ~2003-01 -> 2026-09 (~5,800 sessions, ~23 years). MECHANISM (explicitly NOT 'RSI/Bollinger mean reversion' and NOT the crypto reversal family that died here): short-horizon cross-sectional reversal in equities is the documented compensation for SUPPLYING liquidity to impatient, inventory-driven order flow (Nagel 2012; Campbell-Grossman-Wang). When a liquid large cap sells off hard over one or two sessions with no change in fundamentals, the move is paid for by a dealer/market-maker who must be compensated to hold the inventory; the compensation is realized as a partial bounce over the next one-to-three sessions, and the symmetric effect holds on the up side. This factory has 0/200 on crypto mean reversion and 495 fee_edge deaths — but that is a COST verdict, not a signal verdict: the mechanism pays ~0.2-0.5% per leg per 2-session cycle, which is annihilated by a 0.20% Binance-spot or 0.10% USD-M round trip and survives comfortably on a ZERO-COMMISSION equity venue where the full round-trip friction on these names is ~0.04-0.05% (two ~1bp half-spreads plus impact on sub-$10k clips). The venue choice IS the hypothesis. It is also the structural answer to the overfit/PBO graveyard: unlike the daily ETF regime-rotation sleeves of L99 (35-77 rebalances fit to a handful of macro turning points), this design produces ~125 rebalances/year x 6 positions = ~1,500 trades/year and ~15,000 trades over the joint window, so walk-forward OOS, PBO and holdout are genuinely powered rather than noise. CONSTRUCTION. On every SECOND completed daily bar (a non-overlapping 2-session clock, so positions never stack and the accounting stays trivial): for each of the 12 names compute r2 = the 2-session cumulative close-to-close return, and sigma = the 60-session realized daily return stdev; the ranking score is z = r2 / (sigma * sqrt(2)), i.e. the move measured in that name's OWN volatility units so the ranking is not permanently hijacked by whichever name is the most volatile. NEWS-JUMP EXCLUSION (this is the one economically-motivated filter, not a tuned regime gate): any name with |z| > jump_sigma_cap (default 4.0) is DROPPED from that cycle entirely — moves that large are fundamental repricing (earnings, guidance, M&A, a dividend cut) and the literature is clear that jump-on-news does not revert, while it is exactly where a short leg takes an uncapped gap loss. From the surviving names, go LONG the n_legs=3 lowest z (the oversold/liquidity-demanded side) and SHORT the 3 highest z, equal DOLLAR weights per name, long notional = short notional (dollar-neutral by construction, recomputed from get_account_equity() each cycle), gross exposure 100% of equity => ~16.7% per name, total leverage 1.0x gross within the Reg-T 2x allowance. Hold exactly 2 sessions; flatten everything at the next rebalance bar and re-rank. No pyramiding, no overlapping tranches, no intraday logic (daily bars only — the 1h/1m USEQ history is too short and is deliberately unused). RISK. Per-name hard stop at -8% adverse move from entry on a daily close (caps the residual gap tail the jump filter misses); stopped names stay flat until the next rebalance, and the opposite leg is NOT re-sized (a small transient dollar imbalance is preferable to chasing neutrality with extra trades). Risk per cycle is ~1.5-2% of equity at the stop across the 6 positions. Max exposure per name 20% of equity. No overnight-only or dividend-capture component (dividends are not in PnL on this venue). RUNTIME NOTE FOR THE DEVELOPER (a prior 30-name USEQ cross-sectional strategy here died coding_stuck_3_timeouts, and 12 names x ~5,800 daily bars = ~70k bars): keep every rolling statistic O(1) per bar with fixed-length collections.deque accumulators updated in on_extra_bar/on_bar; do NOT rebuild lists, slice history, or call sorted()/min() over full history inside calculate_signal(). The only sort allowed is over the <=12 current scores, and only on rebalance bars. Extra instruments ride on extra_bar_types; a cycle is skipped (flat) if fewer than 8 names have a fresh bar for the current session, which also handles the session calendar cleanly. PARAMETERS (5, of which 3 are optimizable): lookback_sessions=2, hold_sessions=2 (tied to the rebalance clock), n_legs=3 (fixed by construction), jump_sigma_cap=4.0, stop_loss_pct=8.0 (fixed risk control, not an alpha knob). Optimize only lookback_sessions (1-4), hold_sessions (1-4) and jump_sigma_cap (3.0-6.0) — a deliberately small grid so best-of-N inflation is small. FEE ARITHMETIC (required). These names have a typical daily sigma of ~2.0-2.5%, so a 2-session extreme-decile move is ~4-6%. Documented short-horizon reversal capture on liquid high-idio-vol names is ~10-20% of the prior move over the next 2 sessions => ~0.4-1.0% gross per leg; haircut hard to ~0.25-0.30% per trade for the modern, more-efficient regime and for the fact that the cross-sectional spread is taken on both sides. Against a USEQ round trip of ~0.04-0.05% (zero commission + ~1bp half-spread each way + impact on ~$16k clips in names trading $0.5-3bn/day), that is a 5-6x cost multiple — the single largest reason to run this mechanism on equities and not on any crypto venue.

SMALLEST CHANGE FOR THE REPORTED FAILURE. The gate said: 'metrics unreliable: Incoherent: Sharpe 0.04 disagrees in sign with total_return -10.31% ... Judge on the equity curve.' The sign disagreement itself is the engine's impact-fold aliasing bug I isolated in iteration 5 (runner.py `_sample_points` returns the SAME dicts as mtm['equity_curve'], so `_fold_impact_into_curve` folds 500 of them twice; the dense daily series the Sharpe is computed from therefore carries 2x impact on those points and 1x everywhere else). No strategy code can repair that. The only strategy-side lever over the mis-folded quantity is the size of the modelled impact, so this iteration takes that lever to its end and changes NOTHING else: gross_exposure 0.125 -> 0.05 (10% gross with leverage 2.0). Impact cost scales ~notional^1.5 under the sqrt law, so the $7.41k of impact in the last run (7.4% of starting equity, 12.7x turnover) projects to ~$1.9k, which pulls the equity curve and the mtm curve back together. The calendar clock, marked-equity sizing, price floor, jump filter, whole-share reconciliation and the (lookback 3, hold 4) cell all passed their layers and are untouched, so no earlier layer regresses; the clamp's lower bound moved 0.05 -> 0.02 and is mirrored in _param_bounds. WHY I RECOMMEND ABANDON. Judged on the equity curve as asked, the 2002-2026 record is total_return -10.31%, CAGR -0.44%, max_drawdown 22.1%, profit_factor 0.98, win rate 52.4% over 6,958 trades, avg_trade_return_pct -0.0041% -- an order of magnitude BELOW the 0.05% USEQ floor and on the wrong side of zero. Annual returns are 10 positive / 15 negative with no trend, i.e. a sign-unstable zero mean. Backing out the $7.41k of impact leaves a GROSS (pre-cost) result of about -3% over 24 years: the cross-sectional decile spread this hypothesis rests on is ~0 in the engine, which matches the full-catalog dry-run sweep reported in iteration 4 (registered cell lookback-2/hold-2 = -2.0bp per cycle, t=-0.84; best of all 16 cells +8.7bp at t=1.82, the maximum of the search and not significant). That is a SIGNAL refutation, not a fee verdict -- the whole premise of the hypothesis was that a zero-commission venue would rescue a real premium, and the premium is not there to rescue. Shrinking the book scales a zero-mean P&L toward zero; it cannot create an expectation, and I cannot name a change inside the declared construction or grid that would.

premise_falsified on an adequate, reliable sample. The code implements the hypothesis faithfully and the short-horizon cross-sectional reversal premium it was built to harvest measures ZERO over 6,958 trades and 24 years: mean per-trade return -0.027% at t = -0.41, halves +0.029%/-0.083%, 9 positive years of 25, PF 0.984, Sharpe -0.012, total_return -3.79%. This is a SIGNAL refutation, not a cost verdict — exactly the claim the hypothesis staked itself on. Commissions are $0 and modeled impact only $2,660 on $100k over 24 years, so backing impact out still leaves the gross decile spread at ~-1% cumulative. The per-leg split makes the absence explicit: LONG +0.335%/trade vs SHORT -0.385%/trade is the market's upward drift landing on both sides of a dollar-neutral book, leaving no reversal residual. Optimization cannot help: the shipped cell is the argmax of the developer's own in-sample 16-cell sweep over the declared grid and is still negative, so the whole optimizable space is dead. Not revise_hypothesis: the reframe levers the hypothesis itself declares (zero-commission venue, high-idio-vol large caps, 1-4 session horizon, news-jump screen) are all already applied and the premium is absent on the most favourable venue available. For the Research Lead: the venue-choice argument and friction accounting were sound; what failed is the assumption that a ~0.25-0.30%/leg 2-session liquidity-provision premium still exists on liquid US large caps in 2002-2026.