"ML vs GMM" 是个伪命题 / "ML vs GMM" is a false dichotomy 把争论说成"最大似然 vs GMM"是错误的提法。ML 是 GMM 的特例——它建议一组在明确意义下统计最优的矩。一切都是 GMM；真正的问题是矩的选择：是用辅助统计模型挑出的（哪怕经济上不可解释的）矩，还是用因其经济或数据概括意义而选的（哪怕统计上不有效的）矩。而且根本没有"那个" GMM 估计——GMM 是灵活工具，\(a_T\) 与 \(g_T\) 任你选。二者都是有思想的研究者借以了解数据的工具，而非照搬即得真理的石碑；若不动脑地字面执行，ML 与 GMM 都会给出可怕的结果。（也不必把 GMM 与贴现因子、ML 与期望收益-贝塔强行配对——可对贴现因子用 ML，也可对贝塔模型用 GMM。）Framing the debate as "maximum likelihood vs. GMM" is a bad way to put it. ML is a special case of GMM — it suggests a particular set of moments that are statistically optimal in a well-defined sense. It is all GMM; the real issue is the choice of moments: moments selected by an auxiliary statistical model (even if economically uninterpretable) vs. moments selected for their economic or data-summary interpretation (even if not statistically efficient). And there is no such thing as "the" GMM estimate — GMM is a flexible tool, with any \(a_T\) and \(g_T\) you choose. Both are tools a thoughtful researcher uses to learn what the data say, not stone tablets giving directions that lead to truth if followed literally; followed thoughtlessly, both ML and GMM give horrendous results. (Nor must GMM be paired with the discount factor and ML with expected return-beta — one can do ML on a discount factor, or GMM on a beta model.)

OLS vs GLS：效率与稳健的具体抉择 / OLS vs. GLS: where the trade-off crystallizes GLS 与二阶 GMM 的渐近效率来自协方差/谱密度阵已收敛到总体值；它们用这些矩阵找"测得准"的组合（GLS 找残差方差小的、GMM 找贴现收益方差小的）。危险在于：有限样本里这些量估得差，样本最小方差组合与总体的几无关系。对一个完美模型这不算大问题；但一个很好却不完美的模型，可能很好地给基础组合定价、却很差地给"含强多空头寸"的奇怪线性组合定价（这些头寸因交易/保证金/卖空约束实际不在支付空间内）。故真正的危险是虚假样本最小方差组合与模型设定误差的相互作用。有趣的是，Kandel-Stambaugh (1995)、Roll-Ross (1995) 反而因模型设定误差支持 GLS：只要有任何设定误差（定价误差不恰为零、市场代理不恰在前沿上），就存在能让"期望收益-β 图"任意好或任意坏的组合（取多空 α 即得巨大 α）。GLS 的关键优点是对组合重新打包不变（OLS 的 \(\hat\lambda=(\beta'\beta)^{-1}\beta'E(R^e)\) 依赖打包 \(A\)，GLS 的 \(\hat\lambda=(\beta'\Sigma^{-1}\beta)^{-1}\beta'\Sigma^{-1}E(R^e)\) 不变；HJ 二阶矩权也有此性质）。但这只是事实，不证明 OLS 选的组合好坏——若你不认为 GLS 选的组合有信息量，正可用 OLS 聚焦于经济上有趣的组合。抉择微妙地取决于检验目的：想证伪模型，GLS 帮你聚焦最有力的组合（有效检验本该如此）；但很多模型虽错却"相当好"，丢掉"模型很好地给有趣组合定价"的信息太可惜。明智的折中：报告"有趣"组合上的 OLS 估计，同时报告显示模型被拒的 GLS 检验统计量——这恰是典型的事实集合。GLS and second-stage GMM gain asymptotic efficiency when the covariance/spectral density matrices have converged to population values; they use these to find well-measured portfolios (small residual variance for GLS, small variance of discounted return for GMM). The danger: in finite samples these are poorly estimated, and sample minimum-variance portfolios bear little relation to population ones. For a perfect model this is not a big problem; but a model that is very good yet imperfect may price a basic set of portfolios well while pricing strange long-short combinations badly (positions really outside the payoff space given transactions/margin/short-sale constraints). So the real danger is the interaction of spurious sample minimum-variance portfolios with the model's specification errors. Interestingly, Kandel-Stambaugh (1995) and Roll-Ross (1995) argue for GLS, also from misspecification: so long as there is any misspecification (pricing errors not exactly zero, the market proxy not exactly on the frontier), there exist portfolios producing arbitrarily good or bad expected-return-beta plots (go long positive-α and short negative-α securities for a huge α). GLS's key feature is invariance to repackaging of portfolios (OLS's \(\hat\lambda=(\beta'\beta)^{-1}\beta'E(R^e)\) depends on the repackaging \(A\); GLS's \(\hat\lambda=(\beta'\Sigma^{-1}\beta)^{-1}\beta'\Sigma^{-1}E(R^e)\) does not; the HJ second-moment weight shares this). But this is a fact, not proof that OLS picks good or bad portfolios — if you don't think GLS's choice is informative, use OLS precisely to focus on economically interesting portfolios. The choice subtly depends on the test's purpose: to prove the model wrong, GLS focuses on the most informative portfolios (as an efficient test should); but many models are wrong yet "pretty darn good," and it is a shame to discard the information that the model prices an interesting set of portfolios well. The sensible compromise: report the OLS estimate on "interesting" portfolios, and also report the GLS test statistic showing the model is rejected — which is the typical collection of facts.

GMM 让你评价"被设错"的模型 / GMM lets you evaluate misspecified models ML 对设定误差的回答是："把正确模型写下来再做 ML"——误差相关就建模协方差阵做 GLS，担心代理误差/交易成本/时间聚合/非正态/时变 β 就统统写下来再 ML。可这终究不可行：经济学研究的是定量寓言而非完全设定的模型；我们无法定量描述所有可能的设定误差（若你知道怎么建模它们，它们就不在那儿了）。给 RBC 加"测量误差"以打破奇异性即是例子——但假定的测量误差结构反过来决定了 ML 关注哪些矩，且认真建模测量误差又把我们拉离模型中经济上有趣的部分（既然往往最终归结到合理的矩，何不一开始就直接指定那些矩？）。judiciously 使用的 GMM 让我们把统计火力对准"有趣"的预测、无视世界不符合"无趣"简化的事实：ML 只给 OLS（标准误错）或 GLS（小样本不可信/聚焦无趣处）之选，GMM 则允许你保留 OLS 估计、却为非 i.i.d. 修正标准误；更一般地，允许你指定一组经济上有趣、或你认为对设定误差稳健的矩，而无需说清究竟何种误差使这些矩"最优"。同时 GMM 也灵活地把统计设定误差纳入分布理论（如明知收益非 i.i.d. 正态仍用时序回归——估计不偏，但 ML 公式吐出的标准误不一致，GMM 给出修正）。RBC 的"校准"其实多半就是用经济上合理的矩（产出增长、消费/产出比等）做的 GMM 估计，以避开会使 ML 崩溃的随机奇异性。"judiciously"是重要限定——很多 GMM 估计因随意选矩/资产/工具而受损（如行业组合平均收益几无差异、第七阶滞后几无预测力）。ML's answer to misspecification is "specify the right model, then do ML" — if errors are correlated model the covariance and do GLS; if worried about proxy errors/transaction costs/time aggregation/nonnormality/time-varying β, write them all down and do ML. But this is ultimately infeasible: economics studies quantitative parables, not fully specified models; we cannot quantitatively describe all possible specification errors (if you knew how to model them, they wouldn't be there). Adding "measurement errors" to RBC models to break the singularity is an example — but the assumed error structure then drives which moments ML attends to, and seriously modeling them takes us further from the economically interesting parts (since this often ends up specifying sensible moments anyway, why not specify the sensible moments in the first place?). GMM used judiciously lets us direct the statistical effort at the "interesting" predictions while ignoring that the world does not match the "uninteresting" simplifications: ML offers only OLS (wrong standard errors) or GLS (untrusted in small samples / focused on uninteresting parts), whereas GMM lets you keep an OLS estimate yet correct the standard errors for non-i.i.d.; more generally, it lets you specify an economically interesting or robustly-chosen set of moments without spelling out exactly what misspecification makes those moments "optimal." GMM also flexibly incorporates statistical misspecification into the distribution theory (e.g. knowing returns are not i.i.d. normal but using the time-series regression anyway — the estimate is not inconsistent, but ML's standard errors are; GMM corrects them). RBC "calibration" is often just a GMM parameter estimate using economically sensible moments (output growth, consumption/output ratios) to avoid the stochastic singularity that would doom ML. "Judiciously" is an important qualifier — many GMM estimations suffer from thoughtless choice of moments/assets/instruments (industry portfolios have almost no spread in average returns; the seventh lag predicts little).

"要一个模型才能打败一个模型，而非一次拒绝" / "It takes a model to beat a model, not a rejection" 有说服力、真正改变了人们对数据与模型理解的实证工作，长得很不像教科书里宣讲的统计理论。CAPM 被拒多年后仍被教授、相信、使用——它退场不是因为被拒，而是因为多因子模型提供了另一套自洽的世界观（而多因子模型也被拒！）。即便评价单个模型，最有趣的计算也来自考察具体备择而非笼统的定价误差检验：CAPM 倒下，是因为发现规模、账面市值比等特征进入截面回归，而非因为通用定价误差检验拒绝了它。有影响力的实证工作讲一个故事——若读者看不出是哪些 stylized facts 驱动结果，最有效的程序也说服不了人。聚焦"模型能否解释有趣组合的截面平均收益"的检验，终究比"模型能否解释第二个组合的五阶矩"更有说服力，哪怕 ML 认为后者统计上信息量大得多。Fama-French (1988b, 1993) 即典范：长期可预测性在显著性边缘、多因子模型被 GRS 拒，但它们用平均收益与 β 的表格讲清了驱动结果的稳健事实——统计理论里没有这种表格的位置，但它比一堆 χ² 值有说服力得多。把 t 值从 1.5 改到 2.5 从未实质改变人们对一个问题的看法。统计检验只是评价理论的众多问题之一，且通常不是最重要的——这是对学界如何从一个理论走向另一个理论的实证描述，而非规范主张。Popper/经典统计决策论/Friedman 的"可证伪假设"哲学含一个内在矛盾：要数据说话，方法论作者却不看真实理论如何演化。Kuhn (1970)、McCloskey (1983, 1998) 发现实际过程与形式方法论关系甚微——最大的 t 值并没有获胜。想让自己的想法既正确又有说服力的研究者，宜研究历史上想法是如何说服人的。Empirical work that has been persuasive — that changed people's understanding of the data and which models explain it — looks very different from the statistical theory preached in textbooks. The CAPM was taught, believed, and used for years after formal rejection — it fell not because of a rejection but because the multifactor models offered a coherent alternative worldview (and the multifactor models are also rejected!). Even evaluating a single model, the most interesting calculations come from examining specific alternatives rather than overall pricing-error tests: the CAPM fell when size and book/market were found to enter cross-sectional regressions, not when a generic pricing-error test rejected it. Influential empirical work tells a story — the most efficient procedure does not convince if the reader cannot see what stylized facts drive the result. A test focused on a model's ability to explain the cross-section of average returns of interesting portfolios is ultimately more persuasive than one focused on its ability to explain the fifth moment of the second portfolio, even if ML finds the latter far more informative. Fama-French (1988b, 1993) are exemplars: long-horizon predictability is on the edge of significance and the multifactor model is rejected by GRS, but they made clear, with tables of average returns and betas, what robust facts drive the results — statistical theory has no place for such a table, yet it is far more persuasive than a table of χ² values. Changing a t-statistic from 1.5 to 2.5 has never substantially changed how people think about an issue. Statistical testing is one of many questions in evaluating theories, and usually not the most important — a positive description of how the profession moves from theory to theory, not a normative claim. Popper / classical decision theory / Friedman's "rejectable hypotheses" philosophy contains an inconsistency: it says let the data decide, yet methodologists don't look at how actual theories evolved. Kuhn (1970) and McCloskey (1983, 1998) found the actual process has little to do with formal methodology — the largest t-statistic did not win. A researcher who wants ideas to be convincing as well as right would do well to study how ideas have convinced people in the past.