We extend and recalibrate a modified Chiarella heterogeneous-agent model to allow for arbitrary long-term value drifts, enabling calibration on individual monthly series across indices, bonds, commodities and FX; we find substantial excess volatility (noise-trader–driven price volatility ≈4× fundamental volatility for stock indices) and frequent bimodality of mispricing distributions, and we characterise parameter identifiability via a sloppiness (Fisher/Hessian) analysis that reveals a hierarchy of stiff vs sloppy parameter combinations .
速读内容
Key result — Excess volatility quantified
The ratio Σ = ⟨σ_N / σ_V⟩ (noise trader volatility / fundamental volatility) is large across asset classes, e.g. Indices ≈ 3.87 ± 0.61 (linear) and 3.81 ± 0.65 (non-linear), Commodities ≈ 3.6, FX ≈ 5.9 and Bonds ≈ 13.9, indicating price volatility driven far more by trading noise than by fundamental-value changes .
Interpretation: short-to-medium-term price variance is dominated by σ_N; trend-followers contribute to medium-term decoupling while noise traders amplify short-term volatility .
Data & calibration workflow (EM + Kalman / UKF)
Data: monthly spot prices (Global Financial Data) for indices, commodities, bonds, currencies; long histories (some >200 yrs); various preprocessing (CPI inflation-adjust, remove WWII gaps, de-drifting by polynomial per-decade) .
Calibration steps:
Ex-ante estimate long-term drift G_t (polynomial with ~1 order per decade), detrend price series and estimate trend params α, γ ex-ante .
EM + Kalman smoother on linear model to estimate θ = (κ, β, σ_N, σ_V, v_0) with σ_V tied to σ_N via class-level Σ; for cubic fundamentalist demand use Unscented Kalman Filter (UKF) in E-step .
Multi-step: class-level Σ estimation → per-asset recalibration with σ_V fixed → (if needed) fit cubic κ_3 .
Robustness: dedrifting polynomial order (e.g., k=14,22,30) changes calibrated parameters little → results stable to drift estimation choice .
Trend signal estimation & param choices (quant construction)
Trend signal M_t is EWMA of drift‑adjusted returns; α chosen by maximising Sharpe Ratio of de-drifted trend signal (typical α ≈ 1/5) .
Trend demand functional form: tanh(γ M) justified empirically by return vs trend relationship; γ estimated by parametric fit h(x)=a+b tanh(γ̃ x + c) then rescaled by Var[Ťm] to obtain γ per instrument .
Practical quant implication: trend factor is constructed on de-drifted (mispricing) returns to avoid long‑only bias; signals clipped around ±2σ in practice (saturation) .
Mispricing distribution & bimodality (empirical + numerical)
Empirical Silverman multimodality tests (α=0.02) on filtered vs smoothed mispricings find frequent rejection of unimodality: many indices and several commodities show bimodal mispricing distributions; numerical simulations of the calibrated non-linear model generally confirm bimodality and often show stronger evidence (mitigating small-sample conservatism) .
Jensen-Shannon distances quantify empirical ↔ numerical fit of mispricing PDFs; typical J-S distances are modest (e.g., ≈0.12–0.3), indicating reasonable match despite model simplicity .
Representative calibration statistics (Σ table)
Asset class
Σ (linear) ± sd
Σ (non-linear) ± sd
Indices
3.87 ± 0.61
3.81 ± 0.65
Commodities
3.62 ± 1.62
3.35 ± 1.13
Currencies
5.93 ± 1.10
5.89 ± 1.09
Bonds
13.94 ± 4.46
13.82 ± 4.35
Conclusion: σ_N typically several times larger than σ_V; for indices factor ≈4 consistent with earlier literature (Shiller, Black) .
Fisher/Hessian analysis shows "sloppy" spectrum: a few stiff eigendirections (combinations of κ, σ_N, βγ) are well identified; many directions (e.g., σ_V, α, detailed β vs γ separation) are weakly constrained → large parameter uncertainty in those directions .
Interpretation: key dynamical features (variance of mispricing, bifurcation proximity) are governed by identifiable combinations (e.g., σ_N vs κ and βγ), so model can robustly quantify excess volatility and mispricing modality even if some individual parameters are poorly estimated .
思维导图
深度阅读
以下为对论文 “Revisiting the Excess Volatility Puzzle Through the Lens of the Chiarella Model” 的详尽、逐节解构式分析,覆盖元数据、模型设定、校准、图表解读、估值/方法论分析、风险与局限、斜率(sloppiness)分析到结论性综述。文中所有从原文直接得出的结论或推断在句末均标注了页码溯源标识以便追踪。[注:页码标识采用原文 page_idx 给出的页索引。]
一、元数据与总体概览
标题:Revisiting the Excess Volatility Puzzle Through the Lens of the Chiarella Model。
作者与机构:Jutta G. Kurth, Adam A. Majewski, Jean-Philippe Bouchaud;隶属 École polytechnique / CFM / Capital Fund Management 等机构(详见论文首页作者与脚注)。