Research letters · 30 October 2026
Research letter, 30 October 2026
This week
Two papers this week find that doing less works better: adding a regime overlay and adding return forecasts both lost to the plainer version out of sample, net of costs. The one paper reporting added signal (LLM text for healthcare stock rankings) ran its own placebo. Text shuffled in time within each company predicts about as well as the real text, so most of the "signal" looks like a persistent company fingerprint. The other two items are about clocks and plumbing, not alpha. Everything here is from arXiv. NBER had nothing on topic, and SSRN wasn't searched directly.
1. Market Regime-Based Multi-Asset Portfolio Optimization Using Nested Clustered Optimization for Tactical Asset Allocation
Supriadi, Bambang Siswoyo · v1 20 Oct 2026, v4 29 Oct · arXiv:2610.27751 · code: github.com/obo-cloud/nco-regime-taa
The short version: regime conditioning makes a clustered portfolio worse, through turnover.
Method.
- Regime model: a causal rolling Gaussian HMM with 3 states, chosen by BIC from {2,3,4}. It is fit on a 24-month window using 4 inputs: equity mean and realized vol, bond mean, and the FX return.
- Portfolio: the regime selects the covariance sub-sample fed into Nested Clustered Optimization (NCO).
- Rebalancing: on any regime change, or when weight drift exceeds 5%. Weights apply with a one-period lag.
- Costs: 10 bp on |Δw|.
What they tested and found.
- Universe and sample: 5 Indonesian assets (sovereign bonds, LQ45, gold, CNY, BTC), monthly. Out of sample Jan 2021 – Aug 2025 (56 months).
- NCO + regime: +1.60% cumulative, max drawdown 9.99%, turnover 24.6% a month, 47 rebalances.
- Plain NCO: +29.97%, max drawdown 2.18%, turnover 5.4% a month.
- Significance: Jobson–Korkie/Memmel test of the Sharpe difference, p = 0.007.
- Regime flips: 26 in 56 months. The "bull" state had a negative average equity return out of sample (−1.23% a month), so the labels themselves drifted.
Limitations.
- Small sample: 5 assets and 56 monthly observations. Regime sub-samples often have fewer than 8 observations, so a covariance estimated inside one is close to noise. Much of the result may be "estimating covariance on 6 points is bad", not "regimes are bad".
- Bond returns: the main tables use yield percentage changes as bond returns, and the authors themselves say this is wrong. Their duration-repriced robustness table still favors plain NCO (excess Sharpe +0.14 to +0.26 vs −0.12 to −0.31), which helps.
- No persistence fix tested: no sticky transitions, no confirmation lag. That was the obvious fix, and it's left to future work.
- Deflated Sharpe: mentioned, but no values are reported.
- The paper reached v4 within nine days, and the v4 notes mention "synchronized performance metrics", so earlier versions' numbers may differ from the ones quoted here.
Evidence vs speculation.
- EVIDENCE: in this universe, an unpersisted HMM overlay roughly quintupled turnover and cost most of the return (OOS, net, one market).
- SPECULATION: that persistence constraints would rescue the overlay. This is proposed, not tested.
- SPECULATION: that the result generalizes to larger universes.
What we will do with it. Our own regime classifier uses fixed thresholds with no hysteresis or minimum dwell, on signals standardized over a short window that itself mean-reverts. This paper's failure mode is label churn, and we have never measured ours. We will count label flips per year and median run length over the full regime history, then re-label the same history with one hysteresis variant (wider exit than entry, three consecutive sessions to switch) and score both by the forward 20-session SPY return spread between risk-on and risk-off labels. Hysteresis earns a shadow slot only if it cuts flips by at least 40% and keeps at least 80% of the original spread. If the original label's spread is itself indistinguishable from a block-bootstrap null, we stop tuning thresholds: the classifier has no edge for hysteresis to protect.
2. Simple Portfolios that Deliver Consistent Decent Returns
Oskar Edén Wallberg, Alexander Karolin · 27 Oct 2026 · arXiv:2610.36491 · code: github.com/AlexEmrik/consistent-decent-returns, risk model pip install covforge
Method.
- Universe and trading: long-only, unlevered, monthly portfolios of 59 liquid US ETFs, from 2005-01 to 2025-01.
- Costs and vol: a 55 bp bid-ask cost and an 8% vol target.
- Risk-only methods: min-vol, risk parity and max-diversification, on an EM-fitted statistical factor model.
- Forecast methods: Markowitz with momentum or ridge-regression return forecasts.
- Parameters: fixed at round literature values. The authors say they are "not tuned to the backtest period".
What they tested and found.
- 60/40: Sharpe 0.52, max drawdown 30%.
- Best risk-only portfolios: SPY/TLT/GLD/IBIT max-diversification, Sharpe 0.90, max drawdown 28%; QQQ/XLP/TLT/GLD inverse-vol, Sharpe 0.91.
- Return forecasting: ridge forecasts have negative out-of-sample R² at every penalty. They only approach zero as they shrink to zero. Momentum-Markowitz on the full universe gets Sharpe 0.43. Only after removing whole ETF categories does it reach 0.78, with 81% turnover.
- The authors say plainly that their forecasting attempts "bear no statistical support", and that the momentum gain came from pruning the universe, not from the signal.
Limitations.
- Selection with hindsight: the "best" 3–5 ETF portfolios were picked by scanning every combination from 23 ETFs. By our count that is roughly 44,000 backtests, with no deflated-Sharpe or PBO adjustment. The category ablation that rescued momentum was also chosen on backtest results.
- Survivorship: the universe is ETFs that trade today.
- Inconsistent 60/40 figures: max drawdown is 36% in one section and 30% in another, and volatility is 11% vs 8%.
- So the headline Sharpe numbers for the "best" portfolios are in-sample selections. The negative findings are the credible part, because nobody overfits toward failure.
Evidence vs speculation.
- EVIDENCE: ridge return forecasts on ETFs have negative OOS R² (OOS, 20 years).
- EVIDENCE: diversified risk-only allocation beats 60/40 on Sharpe over 2005–2025. This holds for simple, un-searched versions such as SPY/AGG/GLD equal weight (0.68) and risk parity (0.74).
- SPECULATION: that the top searched combinations will repeat.
What we will do with it. It's the same lesson as our own August review of the production scorer: the forecast layer added nothing, and composition did the work. The useful thing to borrow is a simple, un-searched benchmark. We will add a weekly-rebalanced SPY/AGG/GLD equal-weight chain, fixed in advance, as a reference column in our 2008–2026 replay under the same cost model. If our momentum candidate's net Sharpe does not exceed it, we record that the candidate is not yet beating a no-forecast diversified allocation on a risk-adjusted basis. It is a benchmark, not a signal, so no kill is attached.
3. What Language Models Add to Healthcare-Equity Rankings, and What They Could Have Known
Haotong Fu, Xiyue Zhang, Jack Rose, Ding Sun, Erich Scheller, Eric Rose · 27 Oct 2026 · arXiv:2610.38056 · AI4F workshop, ICAIF '26 · no public code
Method.
- Base model: a random-forest market model on 14 price-volume features, trained through year Y−2 and tested on year Y.
- Added features: screened fundamental and attention features (Newey–West |t| ≥ 4, within-fold Bonferroni) and frozen-embedding text features, blended in with a ridge head.
- Metric: a "neutral IC", the rank IC after regressing out price, dollar volume, vol and Amihud.
- Leakage controls: a frozen pre-2020 GPT-2 rater vs a current model, plus revealed vs blinded identifiers.
What they tested and found.
- Universe and sample: US healthcare names (top half by dollar volume, price ≥ $1), evaluated 2020–2025, excess returns vs XBI.
- 12-month horizon: fundamentals and attention lift neutral IC from 0.063 to 0.093. Text adds about +0.003.
- 3-month horizon: scientific text lifts neutral IC from 0.046 to 0.056 (t 2.8, bootstrap CI [+0.006, +0.015]). The authors label this exploratory.
- Placebo: shuffling text across companies removes the effect. Shuffling it in time within a company predicts as well as the real text, and better in 5 of 6 cells.
- Leakage check: no pre- vs post-cutoff gain for the current model at 1–3 months. The 12-month result is inconclusive because their placebo violates parallel trends.
- Not reported: no portfolio returns or costs.
Limitations.
- Survivorship: the sample is survivor-only. Non-market feeds were pulled in June 2026 for firms listed then.
- Overlap and search: the 12-month claims rest on 65 overlapping months. The text results are unadjusted for the full exploratory search, and some prompt versions weren't logged.
- The placebo is the real result: if within-company shuffled text works, the model is mostly learning which company (a persistent, quasi-static characteristic) rather than what changed. A name-identity baseline would probably capture most of it, and they didn't run one.
Evidence vs speculation.
- EVIDENCE: balance-sheet and attention features carry incremental 12-month neutral IC (OOS by year, survivor-biased).
- SPECULATION: that LLM text carries timely information. Their own placebo argues it is mostly a company fixed effect.
What we will do with it. Healthcare was the worst sleeve in our own recommendation record, which makes the within-name time-shuffle a useful placebo for our scorer. It asks whether ranking skill, where there is any, comes from timely changes or from a sticky per-name level. For each name and date we will replace the score with the same name's score from a random other date at least 60 sessions away, 200 times, and recompute the 20-day rank IC against SPY-excess returns, overall and within Healthcare. If the shuffled IC is at least 80% of the real one, we record that scorer skill is a per-name level effect, and any future scorer work must beat a name-mean baseline, not zero. If the real IC isn't distinguishable from zero in the first place, we stop: there is no skill to decompose.
4. When Does "Post-Call" Begin? An Availability Contract for Earnings-Call NLP Benchmarks
Xiaoyu Ma, Kelvin J. L. Koa, Yuncong Liu, Ke-Wei Huang · 15 Oct 2026 · arXiv:2610.21659 · ICAIF '26 · no repo found
(Posted 15 days before this letter, just outside the usual 14-day window. It's included because it bears directly on our earnings research.)
Method. The authors propose an "Availability Contract" that every event study should declare: the input boundary; evidence that the input was actually available; when the outcome window starts; and which claims that evidence supports. They hold the return endpoint fixed and vary only the start clock: previous close, the call-start bar, or the call-completion bar (reconstructed audio end plus 15 minutes).
What they tested and found.
- Sample: 552 calls from 24 large US firms, 2018–2024, using 5-minute bars.
- Call-start vs completion clock: the raw return's sign flips for 9.4% of calls, and 28.3% of calls change quintile.
- A public benchmark (DEC): shifting before-market targets forward by one session raises 3-day MSE by 18.1–21.4%, while model rankings are unchanged (Kendall τ = 1).
- In short, labels and losses moved, and the leaderboard didn't.
Limitations.
- Proxy clocks: the 15-minute allowance is stipulated and the audio-end time is reconstructed.
- Selective attrition: 1,339 calls were dropped in the timestamp join, which skews the sample toward later, larger firms.
- Intraday only: the paper doesn't address daily-bar work directly.
- Unknown practical impact: the sign-flip numbers are about label definitions, not trading P&L.
Evidence vs speculation.
- EVIDENCE: label definitions are materially clock-sensitive inside events (measured, not a forecast).
- SPECULATION: how much published earnings-NLP alpha disappears under correct clocks.
What we will do with it. Our post-earnings-drift study already declares an event clock: before-open versus after-close timing, unknown timing treated as after-close, and a strictly prior-session regime join. The paper's extra point is to report how sensitive the result is to that clock. We will re-run the headline with every event's entry pushed back one session, and separately with unknown-timing events dropped rather than coerced. If the headline drift shrinks by more than half under the one-session shift, the effect gets labelled first-session-concentrated and timing-fragile before any further work. If its sign flips in either variant, the conclusion is marked unresolved.
5. Where Backtests and Live Trading Part Ways: Six Defects from One Intraday System
Alex Vidovich, Zivana Zerjal · 15 Oct 2026 · arXiv:2610.21846 · also SSRN, DOI 10.2139/ssrn.7517378 · no code
Method. A defect report from a live intraday US equity system, built by replaying live days through the backtest and comparing the two. It is explicitly not a profitability claim.
What they found. None of the six defects showed up in summary statistics:
- Wrong data feed: a data request with no explicit feed parameter returned single-exchange IEX bars instead of consolidated SIP. One stock's 9:30 bar showed about 7,900 shares vs about 3.4M consolidated, roughly 0.2%.
- Dividend-adjusted history: live orders never see adjusted prices.
- Inputs missing from the replay: data the live system downloaded was never requested by the backtest, and missing values silently defaulted.
- Sizing over the account limit: a later size-increasing step skipped a cap. 20 of 61 simulated fills exceeded what the account could afford, and dollar estimates were about 18% too large.
- UTC timestamps read as New York time: a fixed 13:30 UTC key equals 9:30 ET only in summer, which affected 11 scans in winter.
- Settings not reaching the order path: flags that loaded but never reached the code that places orders.
The paper closes with a 10-point checklist: pin the feed and price basis, apply size caps after every sizing step, use exchange-time helpers tested across DST, make missing inputs loud, and test the order path's behavior rather than whether a flag loaded.
Limitations.
- Thin evidence: this is an n=1 anecdote with a 7-day real-money parity record, all in summer time.
- Unnamed providers: the vendor and broker aren't named.
- The paper is useful as a checklist, not as a measurement of how common these defects are.
Evidence vs speculation.
- EVIDENCE: each defect is documented, with its size, in this one system.
- SPECULATION: how common these defects are elsewhere.
What we will do with it. Two of the six map onto questions we can ask of our own pipeline: whether every price loader pins its data feed explicitly, and whether any rolling window mixes adjusted and unadjusted prices. We will audit both. Stored daily volume for a random sample of S&P names gets compared with a consolidated source, and we will count how many names in our breadth universe have a 200-session window on a mixed price basis. A median volume ratio under 0.9, or more than 5% of names on a mixed basis, means a rebuild before any further regime research. Below both thresholds, the audit closes as clean.
Promising experiments
- Within-name time-shuffle placebo on the scorer (item 3). It uses only existing score history, takes about an hour, and gives a sharp answer. It also tells future scorer work which null it has to beat.
- Feed and price-basis audit (item 5). It's cheap and comes first, because if breadth is on a mixed basis, every regime experiment inherits the error.
- Regime label churn and hysteresis (item 1). Run it only after the audit comes back clean. The kill criterion is set so that we don't tune thresholds on a classifier with no forward-return separation.
Watching / not yet credible
- Det-MASO (2610.34936) and LineageEvo (2610.29663): LLM and multi-agent factor mining. They report better IC and Sharpe, but neither abstract mentions how many candidates were searched or any deflation. These are the cases PBO/CSCV exists for. We would wait for code and a held-out era.
- Cross-bond panel learning, convertible bonds (2610.33453): a careful null with purging, fold-specific preprocessing and clustered HAC. MLPs lose to the random walk (holdout RMSE 2.362 vs 2.358, out-of-sample R² −0.004). It's off-universe for us, but a good template for a write-up.
- SAE factors for crypto (2610.39205): lower volatility, but no risk-adjusted edge over a contrarian baseline, and factor identities are unstable across seeds. The authors are honest about it, and it isn't relevant yet.
- Random-feature models and the "SNR frontier" (2610.27255): the idea of a minimum ridge penalty set by signal-to-noise ratio is interesting. The abstract gives no numbers, and the full paper is not yet reviewed.
Updates
Nothing previously covered was materially revised.
This letter is drafted with AI research assistance and reviewed and signed by Bob Katz. It is research commentary, not investment advice. No result described here is a claim net of trading costs unless stated, and the experiments listed are planned tests of our own models, not findings.