Finance & Markets Review — Pilot Issue FMR-2026-06

Forest through the Trees: tree-built cross-sections

穿越树林:基于树构建的横截面

Bryzgalova, Pelger & Zhu (Journal of Finance 2025, E4/F01) use decision trees to endogenously group stocks into interpretable, well-diversified portfolios that span the SDF; over their reported sample, relative to single/double sorts and ML prediction-based portfolios they report up to ~3x higher out-of-sample Sharpe ratios and alphas the authors report as statistically significant. Reported as out-of-sample model performance, not a forward return claim.

P-Trees and the efficient frontier

P-Trees与有效前沿

Cong & Feng (Journal of Financial Economics 2025, E4/F01) grow regression trees on a characteristics panel to build the SDF; Over their reported sample, P-Tree factors attain annualized out-of-sample Sharpe ratios above 3 with alphas the authors report as statistically significant, and test assets yield a GRS statistic of 141.27 against the FF5 model (the authors report this as rejecting the FF5 model at conventional significance). Citation unresolved in our audit — cited by venue+year+title only, no ID attached.

Anomalies vs. short-sale costs

异象 vs. 做空成本

Muravyev, Pearson & Pollet (Journal of Finance 2025, E4/F01): across 162 anomalies the average long-short premium of 0.14%/month (driven by the short leg) falls to -0.01%/month after netting stock-borrow fees, and anomalies are unprofitable even pre-fee once the highest-fee 12% of stock-dates are excluded. A tradability finding, not advice.

Is there a replication crisis? Bayesian verdict

是否存在复制危机?贝叶斯判断

Jensen, Kelly & Pedersen (Journal of Finance 2023, E1/F01) estimate a Bayesian factor-replication model on 153 characteristics in 13 themes across 93 countries and conclude most factors replicate and hold out-of-sample internationally — a meta/replication result, distinct from any single-study claim.

War discourse as a priced factor

战争话语作为定价因子

Hirshleifer, Mai & Pukthuanthong (Journal of Finance 2025, E4/F01) build a war factor from a semi-supervised topic model over ~7M NYT stories across 160 years; it prices the cross-section across traditional and ML test assets spanning 138 anomalies. Reported as in-sample/out-of-sample pricing performance.

Composite securities and price efficiency (theory)

复合证券与价格效率(理论)

Cong, Huang & Xu (NBER 2024, E5/F02) model ETFs and smart-beta products as factor-trading vehicles: proliferation enhances informational price efficiency and raises co-movement, with liquidity effects conditional on factor importance. Theory/model — separated from empirical findings above.

Post-pandemic inflation: a review of the facts

疫情后通胀:事实回顾

Clarida (NBER 2025, E6/F03) identifies three drivers of the 2021-2022 surge (pandemic/Ukraine-war supply shocks, accommodative fiscal-monetary policy, services-to-goods demand shift) and finds advanced-economy central banks reacted slowly then hiked aggressively, achieving substantial disinflation by mid-2024. A review/perspective; supply shocks are highlighted as central to both the rise and the decline.

Are inflation expectations unanchored?

通胀预期是否脱锚?

Coibion & Gorodnichenko (NBER 2025, E6/F03) argue most agents' expectations have been and remain unanchored since 2020; within an expectations-augmented Phillips curve this, plus supply disruptions, accounts for much of the surge and disinflation, and they warn current frameworks are unlikely to re-anchor expectations. Position/perspective.

A generalized Lucas Phillips curve (theory)

广义卢卡斯菲利普斯曲线(理论)

Cochrane (NBER 2025, E5/F03) studies a generalized Lucas Phillips curve where firms learn aggregate demand at varying speeds; unlike standard new-Keynesian models, inflation is initially unstable so a small disinflation builds before reversing, while long-run stability and monetary neutrality are preserved. Theory/model, explicitly separated from empirical claims.

Survey-based r* vs. structural r*

基于调查的r* vs. 结构性r*

Hartley (Mercatus WP 2025, E4/F03) builds model-free survey-based r* estimates for the US, euro area, UK and Canada and finds survey r* has risen since the pandemic while structural Holston-Laubach-Williams estimates fell; the divergence implies very different stance readings (~2-2.5pp restrictive under surveys vs >4-5pp under structural models in 2021-2023). Cited venue+year+title only — no ID attached.

QE/QT and firm financing

QE/QT与企业融资

Eren, Gorea & Zhai (BIS WP 1286, 2025, E4/F03) find that when the Fed conducts QE via government-bond purchases (2011-2024), firms mainly adjust debt maturity, cut interest expense and build cash while total debt, capital and employment stay roughly stable; effects depend on targeted maturities and credit quality, transmitting through bond markets with positive spillovers to high-rated non-US firms. Cited venue+year+title only.

Indirect credit supply via private credit

经由私募信贷的间接信贷供给

Haque, Jang & Wang (FEDS 2025-059, E4/F05) document a bank→BDC→firm intermediation chain: during tightening banks charge BDCs higher rates passed to firms, cushioning credit contraction and supporting investment but strengthening transmission and raising vulnerability (weaker interest-coverage at BDC-dependent firms).

Transformers inside the stochastic discount factor

将Transformer嵌入随机贴现因子

Kelly, Kuznetsov, Malamud & Xu (NBER 2025, E4/F07) embed a transformer in the SDF to exploit conditional pricing via cross-asset information sharing and nonlinearity, reporting large reductions in pricing errors vs prior ML asset-pricing models; a simplified linear-transformer variant clarifies the mechanism. Reported as model performance.

Deep learning with a no-arbitrage criterion

带无套利准则的深度学习

Chen, Pelger & Zhu (Management Science 2024, E4/F07) use deep nets with an adversarial no-arbitrage criterion and extract economic states from many macro series; their individual-stock model outperforms benchmarks out-of-sample on Sharpe ratio, explained variation, and pricing errors. Out-of-sample model performance, not a forward claim.

Can ChatGPT forecast — and for how long?

ChatGPT能预测吗——能持续多久?

Lopez-Lira & Tang (working paper 2025, E4/F07): using post-knowledge-cutoff headlines, GPT-4 captures the initial market response (~90% portfolio-day hit rate for the non-tradable reaction) and its scores predict subsequent drift, especially for small stocks and negative news; forecasting ability rises with model size and strategy returns decline as LLM adoption increases. Past performance is not indicative of future results.

LLM agents in market simulations (testbed)

市场模拟中的LLM智能体(试验平台)

Lopez-Lira (working paper 2025, E5/F07) builds a simulated market (order books, partial fills, dividends) where LLM agents act as value investors, momentum traders or market makers; the markets reproduce price discovery, bubbles, underreaction and strategic liquidity provision — a modeled testbed for LLM effects on stability, not an empirical market claim.

AI as research tool — and as HARKing risk

AI作为研究工具——以及HARKing风险

Eisfeldt & Schubert (NBER 2024, E6/F07) review generative AI's impact on firm valuation and provide a practical introduction for finance researchers; Novy-Marx & Velikov (NBER 2025, E6/F07) mine 30,000+ candidate signals, screen to 96, and use LLMs to auto-generate complete finance papers (including real and fabricated citations), illustrating both efficiency and the danger of industrialized HARKing — directly motivating FMR's citation-audit discipline.

Stablecoins, safe-asset prices and run dynamics

稳定币、安全资产价格与挤兑动态

BIS WP 1270 (2025, E4/F08): stablecoin inflows measurably compress short-term Treasury yields (over the study's estimation sample, roughly a $3.5bn inflow is estimated to lower 3-month T-bill yields ~0.7bp on impact, up to ~4bp within 10 days, with the authors' reported significance/uncertainty), intensifying under stress and as the sector scales. BIS WP 1164 (2025, E5/F08, model): greater reserve transparency is not unambiguously stabilizing — it can raise run risk when reserve confidence is low. Both cited venue+year+title only.

Nonbank intermediation and systemic-risk measurement

非银中介与系统性风险度量

FSB (2025, E6/F08) reports NBFI grew 9.4% in 2024 to $256.8tn (~51% of global financial assets) and maps three bank-NBFI channels while flagging private-credit data gaps. Acharya, Brunnermeier & Pierret (NBER 2024, E4/F08) find market-based systemic-risk measures effective at cross-sectional ranking of vulnerable firms conditional on a stress episode (1895-2023).

Deep learning for fraud detection

用于欺诈检测的深度学习

A systematic review of 57 studies (2019-2024, arXiv 2025, E1/F07) maps deep-learning advances in financial fraud detection, finding CNNs, LSTMs and transformers effective across credit-card transactions, insurance claims and statement audits, while flagging persistent data-imbalance and interpretability challenges.

Crypto trend factor

加密趋势因子

Fieberg, Liedtke, Poddig, Walker & Zaremba (JFQA 2025, E4/F01) construct CTREND, an ML trend factor aggregating price and volume across horizons over 3,000+ coins; over their 3,000+ coin sample it is reported to predict crypto returns with statistically significant loadings, is not subsumed by known factors, and survives transaction costs in large liquid coins. Reported as empirical factor performance, not advice.