Short-Lag Industry-Factor Momentum: Time-Series Sign of Each Sector-minus-SPY Spread's Last-Month Return, SPY-Hedged Long-Short Book (9 SPDR sectors + SPY, USEQ 1-DAY, 21-session rebalance)

Source: Falck et al., 'Is Factor Momentum More than Stock Momentum?' (arXiv 2009.04824, 2020), abstract. The abstract's testable claim is that factor momentum exists both directionally (time-series) and cross-sectionally. It also claims that after controlling for stock momentum, only implementations that include the LAST MONTH of factor returns keep significant Sharpe ratios. Stock-level returns mean-revert at short lags, but factor returns continue at all lags. Our venues hold no anomaly-factor ETFs; the catalog was checked and MTUM/QUAL/USMV/VLUE/IWD/IWF/RSP are all absent. We therefore use the factor class we can build honestly from data on hand: industry factors, each defined as a sector ETF's return minus SPY's. The nine original Select Sector SPDRs (XLK, XLF, XLE, XLV, XLP, XLU, XLI, XLY, XLB) have USEQ 1-DAY bars from Dec-1998, about 27 years. XLRE and XLC are excluded because their history is too short. On every 21st session (counted in bars, not calendar dates), compute each spread's trailing 21-session return, f_i = ln(P_i,t/P_i,t-21) - ln(SPY_t/SPY_t-21). Time-series mode: hold sector i at weight sign(f_i)/9 x gross_fraction, and hold SPY at minus the sum of the sector weights. The SPY leg therefore nets the market out, and the book is a pure bet on industry-factor continuation. Hold for 21 sessions, then recompute. The abstract also tests a cross-sectional variant: long the top 3 and short the bottom 3 spreads by f_i, equal weight, with the SPY leg netting any residual. It is exposed as one categorical switch because both implementations come from the source; there is no other added degree of freedom. The universe is fixed at 10 names, within the ~10-name cap from lesson L173, so per-rebalance compute is trivial (10 log returns). How this differs from the pipeline's LargeSmallCapMonthlyMomentumSpread: that bet is a single size spread. This book is a nine-factor industry book, and the paper's specific claim is the 1-month lag, where stock-level reversal and factor-level continuation pull in opposite directions. What the full paper must confirm: (1) the 1-month formation window and monthly holding period, assumed here from 'last month of returns'; (2) that industry factors behave like their 72 anomaly factors (Moskowitz-Grinblatt 1999 and Arnott et al. 2021 report the same short-lag industry and factor continuation); (3) whether they skip any days between formation and holding (we skip none). Sizing: gross_fraction is FIXED at 0.9 of equity per side, with no leverage tuning and no buying-power parameter. A recent SPY strategy died because its Sharpe changed with position size; this design does not use the margin buffer at all. Net market beta is about 0 by construction. Stop rule: a catastrophic book-level stop flattens everything if the book is down 8% from the last rebalance's equity, and the book re-forms at the next scheduled rebalance.

negative_expectancy / premise_falsified on the first measured report. Code: no_edge (not fee_edge, because the gross edge is itself negative). The test ran 25 years (1999-2024) on 9 SPDR sector-minus-SPY spreads in time-series sign mode. Results: 1478 lots, avg_trade_return_pct -0.351% against a USEQ floor of 0.05% (about 0.10% for a two-sided hedged book), PF 0.959, Sharpe -0.05 with a CI of [-0.41, +0.29], total return -20.3%, max drawdown 34.5% against the pre-registered 20%. Adding back borrow (6.0%) and impact (7.3k) still leaves gross return at about -7%. The executed signal's IC is -0.011 (t = -0.19), so 1-month industry-factor continuation is absent on liquid SPDR sectors. QA's edge concern is borne out and is worse on the full history than on the sandbox's 1653 days (Sharpe 0.18). Benchmark: market-neutral, so buy-and-hold is not comparable. The only untested lever is the categorical cross_sectional mode, which ranks the same 9 spreads. With the time-series signal showing zero IC, it is not worth a 2-hour optimization. If research re-proposes factor momentum, it needs true anomaly-factor returns (for example Ken French portfolios), not sector ETFs.

Iteration 3 fixes the QA findings. (1) Critical: the book stop and rebalance sizing used realized cash balance, which made the stop inert. They now use _mtm_equity(): account balance plus the sum of (last_close - avg_px_open) * signed_qty over every open leg. On a USEQ margin account the balance excludes unrealized PnL, so this is the true book value, and a mark-to-market 8% drawdown within the holding window now triggers the flatten. (2) Warning: when a leg keeps its sign and the size change exceeds $500, _reconcile now submits only the delta (want - have) on the netted position. It closes and reopens only on a sign flip or when the leg goes to zero. This roughly halves turnover and spread/impact cost on the SPY hedge and on unchanged-sign sectors. (3) The calendar-bucket approximation is now documented in the docstring and description. Signal, universe, fixed parameters and tunable domains (mode, lookback_days 15-30) are unchanged. Caveat on the ledger: delta adjustments scale a netted lot, so one closed position can span several rebalances (lesson L175), and per-trade figures describe lots, not rebalances.

Verification failed (Layer 4 — QA review) [class=code_defect]: - [edge_concern] The sandbox smoke test returned +10% over about 1650 sessions, with Sharpe 0.21 and a Sharpe CI of [-0.59, 0.95]. PF was 1.08 on 553 closed lots. Calm and stressed regimes were positive, but the normal regime was negative. impact_cost_pct is 11.5% of gross, and borrow cost is 1.8%. The edge is thin and sensitive to costs. Part of the turnover is self-inflicted by the close-whole-lot-and-reopen resize (see warning). - [critical] The hypothesis's catastrophic book stop (flatten when the book is down 8% from the last rebalance's equity) is effectively inert. It compares self.get_account_equity() with self._rebal_equity. get_account_equity() calls resolve_account_equity(), which returns AccountBalance.total of the USEQ MARGIN account. In NautilusTrader that is the cash balance: it moves only on REALIZED PnL and excludes the unrealized PnL of open positions. This book opens all 10 legs at a rebalance and holds them until the next one, so realized equity barely moves within the holding window. The book could lose 15% mark-to-market and the stop would never fire. Losses are realized only at the next rebalance, and that rebalance resets _rebal_equity and _book_stopped anyway. The hypothesis names this stop explicitly, so a stop that cannot trigger counts as a missing stop. (line 197) — fix: Compute mark-to-market equity: balance plus the sum of unrealized PnL over open positions. Use self.portfolio.unrealized_pnls(venue), or sum position.unrealized_pnl(Price) at each leg's latest aligned close. Use that value both for the 8% stop and for _rebal_equity at the rebalance. Sizing at the rebalance should also use MTM equity; current sizing on realized balance is only correct when the book is flat. - [warning] In _reconcile, when a held leg keeps its sign but its target size changes by more than min_adjust_notional ($500), the code closes the entire lot and reopens the new size. It does not trade the delta. On 0.9/9 = 10% legs this happens at almost every rebalance, so the SPY hedge and the unchanged-sign sectors are round-tripped in full. That roughly doubles turnover and spread/impact cost, and inflates the closed-lot trade count. (line 160) — fix: Submit only the delta (want - have) as a single order on the netted position. Close and reopen only on a sign flip. - [warning] The rebalance bucket is (days_since_epoch*5//7)//21, an approximation of weekdays that is not 'every 21st session counted in bars' as the hypothesis states. This calendar anchoring is acceptable, and is preferred under check 20 for restart and parity stability. Holidays shift the effective cadence by about 1 session per month. Formation (lookback_days rows) and holding (bucket) therefore do not align exactly. That does not change the mechanism. (line 194) — fix: No change is required. Document the approximation in the description.