Search / Rechercher / 搜索ECONOMICS & SOCIAL LAB / AHS RESEARCH
Understand price formation, public-policy trade-offs and their effects on households. Interactive experiments connecting quantitative discipline, the real economy and accountable decision-making.
Explore the experiments ↓An investment cannot be understood through its price alone. Access to energy, fiscal constraints and the distribution of costs also shape its resilience. This laboratory complements the Quant, Trading and Risk Labs by giving scenarios an explicit economic framework.
Models reveal mechanisms, not certainties. All data are synthetic. Assumptions, units, equations and limitations accompany each experiment. Public decisions additionally require observed data, empirical identification and deliberation over collective objectives.
Loading interactive experiments…
The same portfolio can be perceived as a gain or a loss depending on its benchmark. Prospect theory describes relative evaluation; it supplies neither an arbitrage-free price nor a market forecast. Subjective value and decision weights must remain separate from contractual cash flows and pricing probabilities.
Application: present a scenario with a common horizon and currency, then compare interpretations against initial capital, the objective and the benchmark. Synthesis: reframing can change preference without changing payoff; present gains and losses symmetrically.
Assumptions: fix the reference and scale before comparison; distinguish gains from losses; allow diminishing sensitivity. A subjective weight w(p) is not an observed frequency. For multiple outcomes, cumulative prospect theory requires rank-dependent weights; simply summing w(p) is insufficient. A past loss does not mechanically justify taking more risk.
APPLIED RESEARCH · READING PATHWAY
Five original reading notes connect growth, financing, liquidity, rates and debt arithmetic. The scenarios do not describe an observed economy.
What changes when the saving share rises?
Capital, effective labour, saving, depreciation, population and technology growth.
Express output per effective worker as k^α. Compare investment with the capital needed to offset depreciation and effective-labour growth.
A higher saving share changes steady-state capital and transition dynamics; long-run output growth per worker remains driven by exogenous technology.
Requires 0<α<1 and n+g+δ>0. This is a continuous-time reference model, not a country forecast or a policy-effect estimate.
David Romer, Advanced Macroeconomics, 5th ed., McGraw-Hill, 2019, §§1.2–1.3, printed pp. 10–17 / PDF pp. 29–36.
Discuss model review and governanceWhy can weaker borrower wealth reduce investment beyond the direct rate effect?
Borrower net worth, project payoffs, funding need and verification costs.
Map internal funds, external financing and the lender’s information problem. Compare a direct cash-flow shock with a scenario where financing terms also tighten.
The proposed output is a transmission map connecting balance-sheet weakness, agency costs and investment, with the channel stated separately from a calibrated effect.
A conceptual channel does not estimate a credit spread, prove causality or predict a crisis. Contract and institutional assumptions matter.
Romer (2019), §10.2, printed pp. 463–467 / PDF pp. 482–486.
Discuss model review and governanceHow does maturity transformation create fragility, and what spreads it?
Withdrawal rights, liquid resources, liquidation terms and inter-institution links.
Separate early-withdrawal incentives from asset quality. Then document an additional propagation mechanism before moving from an individual institution to system-wide distress.
Use the framework to distinguish liquidity, solvency, common shocks and contagion in a review memorandum.
No run probability or intervention eligibility is calculated. Historical theory is not a statement of current banking law. Systemic propagation is distinct from systematic market exposure.
Romer (2019), §§10.6–10.7, printed pp. 491–492 and 501–502 / PDF pp. 510–511 and 520–521.
Discuss model review and governanceWhich part of a long yield comes from expected short rates?
Policy path scenarios, compounding convention, horizon and term-premium assumption.
Start with the certainty benchmark: a continuously compounded long rate averages the future one-period rates. Under uncertainty, state the expectations and term-premium assumptions explicitly.
Compare short-rate-path and term-premium scenarios before interpreting a duration or curve exposure.
The certainty identity is not an unbiased forecast rule. No live yield curve or estimated policy shock is provided.
Romer (2019), §12.2, printed pp. 583–585 / PDF pp. 602–604.
Discuss model review and governanceWhat primary surplus holds a debt-to-GDP ratio constant?
Initial ratio, annual real rate, real growth, primary surplus and horizon.
Apply the annual recurrence using end-year GDP for the primary balance. Compare the entered surplus with the one-period stabilising surplus.
The demonstrator reports a conditional ratio path and a stabilising balance in percentage points of GDP. A negative ratio represents net assets under the same return assumption.
Constant real rates and growth; no exchange-rate, maturity, stock-flow adjustment or default mechanism. Arithmetic stability is not economic or political feasibility.
Romer (2019), §13.1, printed pp. 662–663 / PDF pp. 681–682. The annual debt-ratio recurrence is an original discrete-time teaching adaptation of the budget identity.
Discuss model review and governanceNo matching note. Clear the search to restore all notes.
Original annual budget-identity scenario. All ratios are percentages of GDP; rate and growth are effective real annual percentages. A positive primary balance is a surplus measured against current-year GDP.
Romer (2019), §13.1, printed pp. 662–663 / PDF pp. 681–682. The annual debt-ratio recurrence is an original discrete-time teaching adaptation of the budget identity.
Reading scope: selected passages and tables of contents were consulted, not the complete books. Notes and numerical examples are original syntheses. The books provide the analytical framework; no author endorsement or independent model validation is implied.
Estimate a relationship, inspect its assumptions, then decide what the evidence supports. The editable examples below are synthetic; they are not economic observations or calibrated forecasts.
y: investment index (points); x: annual financing rate (%); z: confidence index (points). β₁ is measured in investment-index points per percentage point of financing rate; β₂ in investment-index points per confidence-index point. u is the unobserved disturbance, in y units. These labels describe the illustrative dataset.
With an intercept, least squares minimizes squared residuals. Random sampling, appropriate specification, full rank and zero conditional mean are substantive requirements for unbiasedness. Residual orthogonality is a property of the fitted sample; it does not establish exogeneity or causality.
X is the design matrix including the intercept; Xᵢ is its ith row as a column vector. n is the observation count and K the parameter count, including the intercept. HC1 changes covariance estimation, not the coefficients. It allows heteroskedasticity under independent sampling and asymptotic conditions; it does not correct omitted variables or serial dependence.
For equally spaced, correctly ordered observations, the optional Newey–West calculation uses Bartlett weights 1 − l/(L + 1), l = 1,…,L, and the same n/(n − K) finite-sample factor. L is entered, not selected or validated automatically. HAC requires suitable exogeneity, weak dependence and moment conditions. Small samples and persistence can make inference unreliable. No cluster-robust, IV, panel, unit-root or causal estimator is implemented.
01 · Data and specification
03 · Results for calculated inputs
Standard errors and t ratios are descriptive outputs under the selected covariance assumptions. No p-value, confidence interval or automatic causal/significance conclusion is supplied. R² describes in-sample fit, not forecast accuracy.
04 · Charts and residual diagnostics
Charts can be scrolled horizontally on narrow screens. The observation table provides the same numerical information.
05 · Interpretation and variable-omission sensitivity
β̃₁ is the slope from y on x; β̂₁ and β̂₂ come from y on x and z; δ̂₁ comes from z on x. All regressions include an intercept and use the same observations. The relation is an exact sample identity, not an estimate of causal bias. The population omitted-variable-bias interpretation requires the correctly specified model and its exogeneity assumptions.
The intercept is included; K = 2 for simple OLS and K = 3 for multiple OLS. Coefficients are computed by centered/scaled QR decomposition; rank-deficient or numerically near-collinear inputs are rejected. HC1 and HAC use n/(n − K). RMSE is √[SSE/(n − K)], not √(SSE/n). R² is undefined for a constant outcome. Robust errors require more than a finite numerical result.
The auxiliary regression fits squared residuals on the original regressors and an intercept. Its nR² is an asymptotic heteroskedasticity diagnostic with K − 1 regressors; no rejection threshold is applied here. Durbin–Watson is shown only for declared time-ordered rows. It is a descriptive ratio here, not an automatic serial-correlation test. Neither diagnostic identifies omitted factors, simultaneity, measurement error, stationarity or forecast validity.
Jeffrey M. Wooldridge (2020), Introductory Econometrics: A Modern Approach, 7th edition: ch. 2 pp. 23, 42; ch. 3 p. 85, eqs. 3.43–3.46; ch. 8 pp. 265, 270; ch. 12 pp. 399–400, 404.
Wooldridge — publisher reference ↗
Jeffrey S. Zax (2011), Introductory Econometrics: Intuition, Proof, and Practice, Stanford University Press: ch. 4 pp. 89–90; ch. 7 p. 235; ch. 8 p. 296.
H. White (1980), A Heteroskedasticity-Consistent Covariance Matrix Estimator and a Direct Test for Heteroskedasticity, Econometrica 48(4), 817–838. Original covariance reference ↗
W. K. Newey & K. D. West (1987), A Simple, Positive Semi-Definite, Heteroskedasticity and Autocorrelation Consistent Covariance Matrix, Econometrica 55(3), 703–708. Original HAC reference ↗
Continue: identification, instruments and model uncertainty →
Supply parameters to explore the variance response to one innovation. This illustration is not a fitted model, a stochastic simulation or a market forecast. The observation period is your chosen sampling interval; no annualization is applied.
02 · Results for calculated parameters
03 · Conditional-variance forecast
h₀, h₁ and h̄ are variance in squared decimal-return units in the equations. ε₀ and σ₀ are decimal returns; α and β are dimensionless. Displayed variance multiplies h by 10,000 to give %²; displayed volatility multiplies √h by 100. The path table’s √expected variance is a root-mean-square scale, not the expectation of future volatility.
The forecast recursion uses standardized innovations with conditional mean zero and conditional second moment one, fixed parameters and no additional observed innovations. Larger α increases the immediate response to a squared shock; larger β carries more initial variance forward. Equal positive and negative innovations produce the same variance here: asymmetry and leverage effects are not modeled.
The finite-variance restriction makes the long-run variance defined; it does not validate parameters, Gaussian tails, stationarity of an uploaded dataset or suitability for trading. Variance can rise or fall toward its long-run level. No likelihood optimization, calibration, random paths, VaR, expected shortfall or parameter-uncertainty interval is computed.
Chapter basis: Gregoriou & Pascalau (eds.,2011), Sanchis-Marco & Rubia, ch.8 pp.196–197, eq.8.2. Original model: Bollerslev (1986), Generalized autoregressive conditional heteroskedasticity ↗. The conditional-expectation path follows the model recursion; no book dataset or empirical performance is reproduced.
Explore a question, inspect its evidence requirements, then choose a method. The cards are Working papers describing research designs; they do not run estimators. The separate demonstrator below performs only the displayed calculation using synthetic inputs.
Working paper · AHS application
Vintage-dated activity indicators
Filter a latent activity state; report predictive bands.
Revised data must not enter a real-time test.
Source framework; the decision workflow above is an AHS adaptation, not a reproduced case or an empirical finding.
Giordani, Pitt & Kohn, in Geweke, Koop & van Dijk (eds.), The Oxford Handbook of Bayesian Econometrics (2011), ch. 3, printed pp. 61–67 / PDF pp. 74–80.
Reference: supplied book, chapter and printed/PDF pages above.
Working paper · AHS application
Inflation observations and release dates
Use a local-level state-space model to separate signal and measurement noise.
Trend estimates depend on variance assumptions.
Source framework; the decision workflow above is an AHS adaptation, not a reproduced case or an empirical finding.
Giordani, Pitt & Kohn, in Geweke, Koop & van Dijk (eds.), The Oxford Handbook of Bayesian Econometrics (2011), ch. 3, printed pp. 61–67 / PDF pp. 74–80.
Reference: supplied book, chapter and printed/PDF pages above.
Working paper · AHS application
Ordered series with explicit missing values
Propagate the state prediction when a measurement is unavailable.
Missingness must fit the assumed observation process.
Source framework; the decision workflow above is an AHS adaptation, not a reproduced case or an empirical finding.
Giordani, Pitt & Kohn, in Geweke, Koop & van Dijk (eds.), The Oxford Handbook of Bayesian Econometrics (2011), ch. 3, printed pp. 61–67 / PDF pp. 74–80.
Reference: supplied book, chapter and printed/PDF pages above.
Working paper · AHS application
Returns and dated explanatory variables
Specify time-varying regression coefficients and compare filtered sensitivities.
Rapid variation can reflect noise rather than structural change.
Source framework; the decision workflow above is an AHS adaptation, not a reproduced case or an empirical finding.
Giordani, Pitt & Kohn, in Geweke, Koop & van Dijk (eds.), The Oxford Handbook of Bayesian Econometrics (2011), ch. 3, printed pp. 61–67 / PDF pp. 74–80.
Reference: supplied book, chapter and printed/PDF pages above.
Working paper · AHS application
Time series and candidate state count
Use a Markov-switching specification and report state probabilities.
Regime labels and transition persistence need sensitivity tests.
Source framework; the decision workflow above is an AHS adaptation, not a reproduced case or an empirical finding.
Giordani, Pitt & Kohn, in Geweke, Koop & van Dijk (eds.), The Oxford Handbook of Bayesian Econometrics (2011), ch. 3, printed pp. 61–67 / PDF pp. 74–80.
Reference: supplied book, chapter and printed/PDF pages above.
Working paper · AHS application
Identical observations with a fixed information cutoff
Compare current-information filtering with full-sample smoothing.
Smoothed states are not historical trading signals.
Source framework; the decision workflow above is an AHS adaptation, not a reproduced case or an empirical finding.
Giordani, Pitt & Kohn, in Geweke, Koop & van Dijk (eds.), The Oxford Handbook of Bayesian Econometrics (2011), ch. 3, printed pp. 61–67 / PDF pp. 74–80.
Reference: supplied book, chapter and printed/PDF pages above.
Working paper · AHS application
Inflation, activity and rates at one frequency
Estimate a shrinkage BVAR and compare predictive densities with a simple benchmark.
Prior tightness and sample length affect results.
Source framework; the decision workflow above is an AHS adaptation, not a reproduced case or an empirical finding.
Del Negro & Schorfheide, in The Oxford Handbook of Bayesian Econometrics (2011), ch. 7, printed pp. 293–305 / PDF pp. 306–318.
Reference: supplied book, chapter and printed/PDF pages above.
Working paper · AHS application
Rates, output, prices and an identification scheme
Compute structural impulse responses under explicit restrictions.
A reduced-form correlation does not identify a policy effect.
Source framework; the decision workflow above is an AHS adaptation, not a reproduced case or an empirical finding.
Del Negro & Schorfheide, in The Oxford Handbook of Bayesian Econometrics (2011), ch. 7, printed pp. 293–305 / PDF pp. 306–318.
Reference: supplied book, chapter and printed/PDF pages above.
Working paper · AHS application
Nonstationary macro series with aligned definitions
Compare cointegrated VAR specifications and posterior long-run uncertainty.
A common trend alone does not establish cointegration.
Source framework; the decision workflow above is an AHS adaptation, not a reproduced case or an empirical finding.
Del Negro & Schorfheide, in The Oxford Handbook of Bayesian Econometrics (2011), ch. 7, printed pp. 293–305 / PDF pp. 306–318.
Reference: supplied book, chapter and printed/PDF pages above.
Working paper · AHS application
Country panels, rates and trade indicators
Use a multicountry Bayesian system to trace conditional transmission.
Identification and cross-country comparability remain necessary.
Source framework; the decision workflow above is an AHS adaptation, not a reproduced case or an empirical finding.
Del Negro & Schorfheide, in The Oxford Handbook of Bayesian Econometrics (2011), ch. 7, printed pp. 293–305 / PDF pp. 306–318.
Reference: supplied book, chapter and printed/PDF pages above.
Working paper · AHS application
Macro observations and an explicit structural model
Compare DSGE policy scenarios integrating parameter uncertainty.
Results inherit equilibrium assumptions and policy credibility conditions.
Source framework; the decision workflow above is an AHS adaptation, not a reproduced case or an empirical finding.
Del Negro & Schorfheide, in The Oxford Handbook of Bayesian Econometrics (2011), ch. 7, printed pp. 293–305 / PDF pp. 306–318.
Reference: supplied book, chapter and printed/PDF pages above.
Working paper · AHS application
Candidate models and common forecast targets
Combine predictive densities using posterior model probabilities.
Model probabilities are conditional on the candidate set and priors.
Source framework; the decision workflow above is an AHS adaptation, not a reproduced case or an empirical finding.
Del Negro & Schorfheide, in The Oxford Handbook of Bayesian Econometrics (2011), ch. 7, printed pp. 293–305 / PDF pp. 306–318.
Reference: supplied book, chapter and printed/PDF pages above.
Working paper · AHS application
Ordered asset returns
Infer latent log variance and its posterior uncertainty.
Latent volatility is model-dependent, not directly observed.
Source framework; the decision workflow above is an AHS adaptation, not a reproduced case or an empirical finding.
Jacquier & Polson, in The Oxford Handbook of Bayesian Econometrics (2011), ch. 9, printed pp. 439–445 / PDF pp. 452–458; stochastic-volatility framework also ch. 3, pp. 63–64 / PDF pp. 76–77.
Reference: supplied book, chapter and printed/PDF pages above.
Working paper · AHS application
Return shocks and a latent variance process
Allow correlation between return and volatility innovations.
Timing conventions must be consistent in estimation and forecasting.
Source framework; the decision workflow above is an AHS adaptation, not a reproduced case or an empirical finding.
Jacquier & Polson, in The Oxford Handbook of Bayesian Econometrics (2011), ch. 9, printed pp. 439–445 / PDF pp. 452–458; stochastic-volatility framework also ch. 3, pp. 63–64 / PDF pp. 76–77.
Reference: supplied book, chapter and printed/PDF pages above.
Working paper · AHS application
Return series and distributional alternatives
Compare predictive tail probabilities under Gaussian and fat-tailed specifications.
Sparse extremes make tail conclusions uncertain.
Source framework; the decision workflow above is an AHS adaptation, not a reproduced case or an empirical finding.
Jacquier & Polson, in The Oxford Handbook of Bayesian Econometrics (2011), ch. 9, printed pp. 439–445 / PDF pp. 452–458; stochastic-volatility framework also ch. 3, pp. 63–64 / PDF pp. 76–77.
Reference: supplied book, chapter and printed/PDF pages above.
Working paper · AHS application
Asset returns, constraints and risk preferences
Optimize predictive utility after integrating uncertain return parameters.
This is a research design, not an investment recommendation.
Source framework; the decision workflow above is an AHS adaptation, not a reproduced case or an empirical finding.
Jacquier & Polson, in The Oxford Handbook of Bayesian Econometrics (2011), ch. 9, printed pp. 439–445 / PDF pp. 452–458; stochastic-volatility framework also ch. 3, pp. 63–64 / PDF pp. 76–77.
Reference: supplied book, chapter and printed/PDF pages above.
Working paper · AHS application
Return histories and a declared prior
Compare unregularized and empirical-Bayes expected-return estimates.
Reusing data to calibrate priors needs explicit disclosure.
Source framework; the decision workflow above is an AHS adaptation, not a reproduced case or an empirical finding.
Jacquier & Polson, in The Oxford Handbook of Bayesian Econometrics (2011), ch. 9, printed pp. 439–445 / PDF pp. 452–458; stochastic-volatility framework also ch. 3, pp. 63–64 / PDF pp. 76–77.
Reference: supplied book, chapter and printed/PDF pages above.
Working paper · AHS application
Aligned returns with inception dates
Use a joint likelihood to exploit longer and shorter histories together.
Missing-history and dependence assumptions determine information transfer.
Source framework; the decision workflow above is an AHS adaptation, not a reproduced case or an empirical finding.
Jacquier & Polson, in The Oxford Handbook of Bayesian Econometrics (2011), ch. 9, printed pp. 439–445 / PDF pp. 452–458; stochastic-volatility framework also ch. 3, pp. 63–64 / PDF pp. 76–77.
Reference: supplied book, chapter and printed/PDF pages above.
Working paper · AHS application
Fund returns, benchmark and fees
Infer alpha conditional on the stated factor model.
Omitted factors and selection bias can mimic skill.
Source framework; the decision workflow above is an AHS adaptation, not a reproduced case or an empirical finding.
Jacquier & Polson, in The Oxford Handbook of Bayesian Econometrics (2011), ch. 9, printed pp. 439–445 / PDF pp. 452–458; stochastic-volatility framework also ch. 3, pp. 63–64 / PDF pp. 76–77.
Reference: supplied book, chapter and printed/PDF pages above.
Working paper · AHS application
Matched asset and hedge returns
Transform joint posterior draws into hedge-ratio uncertainty.
Basis risk and transaction costs remain outside a simple covariance ratio.
Source framework; the decision workflow above is an AHS adaptation, not a reproduced case or an empirical finding.
Jacquier & Polson, in The Oxford Handbook of Bayesian Econometrics (2011), ch. 9, printed pp. 439–445 / PDF pp. 452–458; stochastic-volatility framework also ch. 3, pp. 63–64 / PDF pp. 76–77.
Reference: supplied book, chapter and printed/PDF pages above.
Working paper · AHS application
One dataset and several defensible priors
Compare posterior and predictive changes before selecting a specification.
A narrow posterior can simply reflect a strong prior.
Source framework; the decision workflow above is an AHS adaptation, not a reproduced case or an empirical finding.
Giordani, Pitt & Kohn, in Geweke, Koop & van Dijk (eds.), The Oxford Handbook of Bayesian Econometrics (2011), ch. 3, printed pp. 61–67 / PDF pp. 74–80.
Reference: supplied book, chapter and printed/PDF pages above.
Working paper · AHS application
Time-ordered forecast densities and later outcomes
Count out-of-sample tail events without refitting on future observations.
Coverage alone does not establish independence or economic usefulness.
Source framework; the decision workflow above is an AHS adaptation, not a reproduced case or an empirical finding.
Jacquier & Polson, in The Oxford Handbook of Bayesian Econometrics (2011), ch. 9, printed pp. 439–445 / PDF pp. 452–458; stochastic-volatility framework also ch. 3, pp. 63–64 / PDF pp. 76–77.
Reference: supplied book, chapter and printed/PDF pages above.
Working paper · AHS application
Posterior parameter draws and future innovation draws
Compare fixed-parameter forecasts with fully predictive distributions.
The comparison is conditional on the chosen model.
Source framework; the decision workflow above is an AHS adaptation, not a reproduced case or an empirical finding.
Del Negro & Schorfheide, in The Oxford Handbook of Bayesian Econometrics (2011), ch. 7, printed pp. 293–305 / PDF pp. 306–318.
Reference: supplied book, chapter and printed/PDF pages above.
Working paper · AHS application
Alternative admissible restrictions
Recompute policy responses across identification assumptions.
A proper posterior does not resolve weak identification.
Source framework; the decision workflow above is an AHS adaptation, not a reproduced case or an empirical finding.
Del Negro & Schorfheide, in The Oxford Handbook of Bayesian Econometrics (2011), ch. 7, printed pp. 293–305 / PDF pp. 306–318.
Reference: supplied book, chapter and printed/PDF pages above.
Interactive demonstrator · AHS · 15 October 2026
AHS analytical illustration of Bayesian updating: independent normal observations, unknown mean and known observation standard deviation. User inputs are synthetic sufficient statistics; no dataset is fitted.
m₀, s₀: prior mean and standard deviation; n: sample size; ȳ: sample mean; σ: known observation standard deviation; mₙ, vₙ: posterior mean and variance; s_pred: predictive standard deviation for one new observation. Means and standard deviations use percentage points.
Giordani, Pitt & Kohn, in Geweke, Koop & van Dijk (eds.), The Oxford Handbook of Bayesian Econometrics (2011), ch. 3, printed pp. 61–67 / PDF pp. 74–80.
The equation and implementation are simplified AHS applications of the cited framework.
APPLIED RESEARCH · READING PATHWAY
These extensions deepen the existing econometrics and Bayesian cases. A smoother is not a causal estimator; a conditional filter is not a complete parameter posterior.
How much smoothing is supported by the sample?
Seven fixed synthetic x values, seven editable responses, bandwidth and query location.
Compare Gaussian local-constant fits at several bandwidths. Leave each point out before predicting it; report the average squared held-out error.
The chart shows the entered points and their fitted conditional-mean curve. Its LOOCV score compares bandwidth choices for this teaching sample only.
Not an additive-model estimator, causal effect or prediction interval. Leave-one-out validation assumes an exchangeable sample; dependent financial time series need a different validation design.
Shujie Ma & Lijian Yang, ch. 6, §§6.1–6.2, printed pp. 149–154 / PDF pp. 169–174, in Racine, Su & Ullah (eds.), The Oxford Handbook of Applied Nonparametric and Semiparametric Econometrics and Statistics, Oxford University Press, 2014. The one-dimensional Gaussian local-constant demonstration is an original reduction, not the chapter’s two-step additive estimator.
Discuss model review and governanceCan nonlinear drivers be read separately without estimating a full high-dimensional surface?
Response, explanatory variables, centring convention and candidate interactions.
Specify an intercept plus centred univariate functions. Check whether material interactions contradict additivity before interpreting components as separate effects.
A proposed review compares additive, linear and interaction-rich alternatives using the same held-out target.
Additivity reduces dimensionality by assumption, not by proving interactions absent. Correlated predictors can make component interpretation unstable.
Joel L. Horowitz, ch. 5, §5.1, printed pp. 129–131 / PDF pp. 149–151, in Racine, Su & Ullah (eds.), 2014.
Discuss model review and governanceWhen does another basis function add information rather than fit noise?
Common target, candidate polynomial or spline bases, and held-out folds.
Estimate finite basis expansions at several orders. Select or average using errors computed on observations excluded from fitting; keep preprocessing inside each training fold.
Report error versus basis dimension and a sensitivity table; separate selection of one model from weighted averaging of several.
The cited asymptotic results require their regularity assumptions. For time series, chronological validation is a proposed adaptation, not a guarantee supplied by an iid theorem.
Bruce E. Hansen, ch. 8, introduction and §8.2, printed pp. 215–217 / PDF pp. 235–237, in Racine, Su & Ullah (eds.), 2014.
Discuss model review and governanceWhy can small data errors create large changes in an estimated function?
Moment operator, estimated target, noise scale and regularisation choices.
State the inverse problem before fitting. Examine sensitivity to perturbations and regularisation strength rather than treating an exact numerical fit as identification evidence.
For a proposed inverse-calibration review, preserve the operator, penalty and sensitivity path. This financial application is an editorial analogy to the chapter’s inverse-moment setting.
A numerical penalty is not a proof of economic identification. No nonparametric instrumental-variable estimator is implemented here.
Marine Carrasco, Jean-Pierre Florens & Eric Renault, ch. 3, §3.1, printed pp. 65–66 / PDF pp. 85–86, in Racine, Su & Ullah (eds.), 2014.
Discuss model review and governanceDoes a good nonlinear fit survive the series’ trend structure?
Time-ordered levels, transformations, integration assumptions and break dates.
Declare which series are stationary or integrated and why. Compare economically justified level relations with transformations; use tests designed for the stated dependence and trend regime.
The proposed output is a specification record showing why smoothing was applied to levels, differences or residuals.
Differencing is not an automatic cure. A nonlinear transformation of an integrated process need not retain the same integration classification.
Yiguo Sun & Qi Li, ch. 14, §14.1, printed pp. 444–445 / PDF pp. 464–465, in Racine, Su & Ullah (eds.), 2014.
Discuss model review and governanceWhat uncertainty does a local-level filter actually report?
Prior mean and variance, process variance Q, observation variance R and ordered observations.
Predict the latent state, compute the gain and update after each observation. Keep the variances fixed so the displayed uncertainty is explicitly conditional on these inputs.
The new demonstrator turns the existing state-space case into a reproducible filter. State intervals differ from next-observation prediction intervals and from full-sample smoothing.
Gaussian independent errors, scalar random walk and known Q/R. No parameter posterior, regime inference, missing-data treatment or real-time macro nowcast is estimated.
Paolo Giordani, Michael Pitt & Robert Kohn, ch. 3, §2.2, printed pp. 64–66 / PDF pp. 77–79, in Geweke, Koop & van Dijk (eds.), The Oxford Handbook of Bayesian Econometrics, Oxford University Press, 2011. The scalar known-variance filter is an original restricted implementation.
Discuss model review and governanceHow much information remains when simulated draws are correlated?
Post-warmup draws of a stated posterior functional, trace and batching choices.
Estimate the numerical standard error using dependence-aware methods such as batch means. Compare it with the independent-draw variance to obtain an inefficiency factor and functional-specific ESS.
A sampling review should preserve the target, initialisation, mixing evidence and numerical error for each reported quantity.
ESS is not a convergence certificate and is not a universal property of an entire chain. This note does not implement a general MCMC sampler.
Siddhartha Chib, ch. 5, §§3.3–3.4, printed pp. 193–195 / PDF pp. 206–208, in Geweke, Koop & van Dijk (eds.), 2011.
Discuss model review and governanceNo matching note. Clear the search to restore all notes.
Original one-dimensional local-constant smoother. Fixed synthetic x values: −3, −2, −1, 0, 1, 2, 3 (dimensionless). Enter seven y values in arbitrary units. The LOOCV score is in squared y units; no prediction band is estimated.
Shujie Ma & Lijian Yang, ch. 6, §§6.1–6.2, printed pp. 149–154 / PDF pp. 169–174, in Racine, Su & Ullah (eds.), The Oxford Handbook of Applied Nonparametric and Semiparametric Econometrics and Statistics, Oxford University Press, 2014. The one-dimensional Gaussian local-constant demonstration is an original reduction, not the chapter’s two-step additive estimator.
Original scalar Gaussian filter with known variances: xₜ=xₜ₋₁+ηₜ, yₜ=xₜ+εₜ. The first observation follows one transition from prior state x₀. Q and R are variances in squared observation units, not standard deviations.
Paolo Giordani, Michael Pitt & Robert Kohn, ch. 3, §2.2, printed pp. 64–66 / PDF pp. 77–79, in Geweke, Koop & van Dijk (eds.), The Oxford Handbook of Bayesian Econometrics, Oxford University Press, 2011. The scalar known-variance filter is an original restricted implementation.
Reading scope: selected passages and tables of contents were consulted, not the complete books. Notes and numerical examples are original syntheses. The books provide the analytical framework; no author endorsement or independent model validation is implied.
Choose the estimand before choosing the estimator: mean, conditional variance, dependence or tail risk. These six reading modules extend the regression workbench with model frameworks and validation questions. They do not add automated GARCH, Kalman, factor or quantile estimators.
An apparently stable return average can coexist with changing risk. GARCH models let yesterday’s innovation and variance influence today’s conditional variance. This is useful for comparing a calm period with a volatility cluster, rather than applying one unconditional standard deviation to every date. A model for the mean must be specified first; predictable movements should not automatically be treated as volatility shocks.
rₜ and μₜ: return and conditional mean, in decimal return per observation period. εₜ: mean innovation in the same unit. hₜ and ω: squared-return units; zₜ, α and β: dimensionless. Standardized innovations have conditional mean zero and variance one.
Use ω > 0 and α, β ≥ 0 for positivity. In the standard finite-variance specification, α + β < 1 gives the long-run variance ω/(1 − α − β). Near-unit persistence makes horizon and initialization important. These conditions do not imply Gaussian tails. Compare standardized residuals and their squares, consider asymmetry and heavy-tailed alternatives, and evaluate forecasts on later observations. Robust OLS errors do not estimate this conditional variance process.
Gregoriou & Pascalau (eds., 2011), Sanchis-Marco & Rubia, ch. 8, pp. 196–197, eq. 8.2; Terrell & Fomby (eds., 2006), Baur, pp. 6–10.
Changing asset volatilities and changing correlations are different sources of portfolio risk. A constant-correlation model can still produce time-varying covariance. Multivariate GARCH separates these effects and makes the covariance matrix, rather than isolated pairwise estimates, the object to validate. For a diversified portfolio, compare a constant-correlation benchmark with a dynamic specification over the same sample.
Hₜ: conditional return covariance matrix (squared-return units). Dₜ: diagonal conditional standard deviations (return units). Rₜ: correlation matrix; w: portfolio weights, both dimensionless. The second equation is a BEKK-type convention: C has return units, A and B are dimensionless, and ε is the return-innovation vector.
With CᵀC positive definite and admissible initialization, the BEKK recursion preserves positive definiteness. This alone does not establish stationarity or a good forecast. DCC-type approaches estimate marginal variances before dependence, while Baur’s flexible dynamic correlation model is a distinct bivariate proposal. Their parameters and restrictions should not be interchanged. Validate symmetry, eigenvalues, residual cross-dependence, estimation stability and out-of-sample portfolio variance. Statistical spillovers do not establish economic causality.
Terrell & Fomby (eds., 2006), Dirk Baur, A Flexible Dynamic Correlation Model, pp. 3–10, especially eq. 4; Gregoriou & Pascalau (eds., 2011), Mancino & Sanfelici, ch. 1, pp. 3–9.
Robert F. Engle & Kenneth F. Kroner (1995). Multivariate Simultaneous Generalized ARCH. Econometric Theory 11(1), 122–150.
Foundational BEKK parameterization and covariance restrictions ↗Robert Engle (2002). Dynamic Conditional Correlation: A Simple Class of Multivariate Generalized Autoregressive Conditional Heteroskedasticity Models. Journal of Business & Economic Statistics 20(3), 339–350.
Original DCC publication ↗BEKK and DCC are distinct parameterizations. These links document the covariance and dynamic-correlation methods discussed above; they do not turn the portfolio calculator into a fitted multivariate GARCH estimator. Baur’s bivariate FDC contribution remains attributed to its book chapter.
More frequent sampling does not automatically produce a better risk estimate. Bid–ask bounce, price discreteness and asynchronous transactions contaminate observed prices. Realized variance aggregates squared log-price changes, but at very fine sampling intervals those changes can be dominated by observation noise. A useful study compares sampling frequencies and explicitly records how missing or asynchronous observations are handled.
Yᵢ: observed log price; pᵢ: latent efficient log price; uᵢ: observation noise, all dimensionless with a fixed price-unit convention. RV: squared log-return units over the stated window; m: number of intervals. RV is variance, whereas √RV is volatility; annualization requires an explicit time convention.
Jungbacker and Koopman formulate noisy prices as a state-space problem; Kalman filtering applies to the linear Gaussian case, with more elaborate volatility specifications requiring additional estimation methods. Mancino and Sanfelici discuss Fourier covariance estimation and asynchronous sampling. This is a statistical covariance method, distinct from Fourier option pricing. Noise assumptions, timestamp alignment, sampling window and positive semidefiniteness must be checked. Do not interpolate missing prices silently or interpret microstructure-induced return autocorrelation as market inefficiency.
Terrell & Fomby (eds., 2006), Jungbacker & Koopman, pp. 186–190, eqs. 4–7; Gregoriou & Pascalau (eds., 2011), Mancino & Sanfelici, ch. 1, pp. 4–9.
Factor regression asks which common return components explain an asset or portfolio. The three-factor specification adds size and value portfolios to market excess return. It is useful for distinguishing a strategy’s common exposures from its unexplained intercept. An estimated intercept is conditional on the factors, return construction and observation period; it is not proof of manager skill or a promised excess return.
All returns and factors use decimal returns for the same period and currency. SMB is small-minus-big portfolio return; HML is high-minus-low book-to-market portfolio return. aᵢ and εᵢ have return units; bᵢ, sᵢ and vᵢ are dimensionless loadings. The symbol vᵢ avoids confusing a value loading with conditional variance hₜ.
Align dates and risk-free rates; control survivorship and look-ahead bias; specify rebalancing, dividends and gross versus net returns. Check rank, residual dependence and coefficient stability. A high in-sample R² does not validate asset-pricing restrictions or forecasting power. The current workbench accepts one or two explanatory columns with fixed economic labels; it does not implement this three-factor dataset or estimate all three loadings. Use the existing Portfolio tools for allocation questions, not as a substitute for factor estimation.
Gregoriou & Pascalau (eds., 2011), Allen, Singh & Powell, ch. 7, pp. 176–180, eq. 7.3.
OLS estimates a conditional mean. Quantile regression instead studies a chosen part of the conditional distribution, allowing market exposure to differ between lower-tail, central and upper-tail outcomes. For example, compare the financing or market exposure associated with weak outcomes against the exposure near the median. Quantile coefficients describe conditional distributional associations; they do not trace the causal response of the same individual observation across quantiles.
τ: dimensionless quantile level. yᵢ and residual u have outcome units. Xᵢ includes an intercept and covariates; each component of bτ has outcome units per corresponding covariate unit. The indicator is one for negative residuals and zero otherwise. At τ = 0.5, the objective is proportional to absolute-error loss.
Tail estimates need enough relevant observations and appropriate dependence-aware inference. Check specification, influential covariates, quantile crossing and stability across samples. Robustness to large outcome residuals does not eliminate leverage problems or endogeneity. Median regression differs from OLS unless additional distributional conditions apply. HC1 or Newey–West errors around an OLS fit do not turn it into a quantile regression. This section explains the method; the existing calculator still estimates OLS.
Gregoriou & Pascalau (eds., 2011), Allen, Singh & Powell, ch. 7, pp. 178–180, eq. 7.4.
A volatility estimate, a downside moment and a return quantile answer different questions. A lower partial moment measures shortfall relative to a chosen target. A conditional lower return quantile supplies a VaR threshold under a declared sign convention. Liquidity and trading variables may add predictive information, but their relevance must be tested on observations unavailable during estimation.
r and target c: decimal returns over the same horizon; LPMₖ has return units to power k. τ: lower-tail probability. Qτ and the return-based VaR are in return units; positive VaR denotes a loss threshold, but can be negative if the relevant return quantile is positive. For a fixed linear position of value V, monetary VaR is V times return VaR. Iₜ is a dimensionless exception indicator.
Choose horizon, target and return convention before estimation. For continuous returns, a well-specified quantile targets exception frequency τ; frequency alone is insufficient because clustered exceptions can reveal conditional failure. Use chronological training and validation windows, inspect coverage and exception dependence, compare forecast loss and document refitting rules. Do not randomly shuffle financial time series or select a model repeatedly on the final test sample. The cited chapters contain historical applications, not current regulatory requirements or evidence of this site’s investment performance.
Gregoriou & Pascalau (eds., 2011), Iqbal, Brooks & Galagedera, ch. 6, pp. 154–158; Sanchis-Marco & Rubia, ch. 8, pp. 194–198, 204–206.
Robert F. Engle & Simone Manganelli (2004). CAViaR: Conditional Autoregressive Value at Risk by Regression Quantiles. Journal of Business & Economic Statistics 22(4), 367–381.
Original CAViaR publication ↗Peter F. Christoffersen (1998). Evaluating Interval Forecasts. International Economic Review 39(4), 841–862.
Original conditional-coverage evaluation reference ↗CAViaR models the evolution of a conditional quantile; conditional-coverage evaluation asks whether forecast exceptions have the intended frequency and dependence properties. They answer different questions. These publications support the methodological reading, not a claim that the site estimates CAViaR or runs the corresponding likelihood-ratio tests.
Related tools: the regression workbench above estimates means and covariance of coefficients. The Portfolio and Risk Labs explore allocation and exposure under their own stated assumptions; they do not calibrate the models described here.
Dek Terrell & Thomas B. Fomby (eds., 2006). Econometric Analysis of Financial and Economic Time Series, Advances in Econometrics, vol. 20, part A. Elsevier. ISBN 978-0-7623-1274-0. Chapter authors and page ranges are recorded in each module.
Greg N. Gregoriou & Razvan Pascalau (eds., 2011). Financial Econometrics Modeling: Market Microstructure, Factor Models and Financial Risk Measures. Palgrave Macmillan. ISBN 978-0-230-28362-6.
The explanatory text is an original synthesis of the cited chapters. Formulas use explicitly declared notation; no empirical tables, performance results or book illustrations are reproduced.
Start with the quantity the decision needs. Better coefficient standard errors, a more realistic volatility process and a better tail forecast solve different problems. The table separates the available calculator from the research methods described in the reading modules.
| Question | Method | Required data | Decisive checks | Available here |
|---|---|---|---|---|
| Explain an average relationship | OLS + appropriate coefficient covariance | Outcome and aligned explanatory variables | Specification, exogeneity, rank and residual dependence | Interactive workbench above |
| Forecast changing variance | GARCH / stochastic volatility | Ordered returns and a specified mean model | Persistence, innovation distribution and future forecast loss | Reading framework + supplied-parameter GARCH scenario |
| Measure joint portfolio risk | Multivariate covariance / dynamic correlation | Aligned multi-asset returns and portfolio weights | Positive semidefiniteness, stability and realized portfolio risk | Reading framework; existing allocation tools linked |
| Measure intraday variation | Noise-aware realized or state-space methods | Timestamped trades/quotes and observation conventions | Noise, asynchronous sampling and sensitivity to frequency | Reading framework |
| Explain distributional exposure | Factor models / quantile regression | Matched returns and factors; declared quantile level | Factor construction, leverage, tail sample size and coefficient stability | Reading framework |
| Validate a loss threshold | Chronological quantile backtesting | Forecasts made before outcomes; fixed horizon and units | Coverage, clustered exceptions and quantile forecast loss | Reading framework; Risk Lab linked |
These are proposed analytical workflows, not executed empirical studies. No dataset, fitted parameter or performance result is attributed to Alpha Hedge Services.
Question: does a return series reflect common factor exposure or a residual intercept? Use total returns, a matching risk-free series and factors with the same currency and period. Separate fee treatment and trading costs from statistical attribution. Begin with a market-only benchmark, then compare a documented factor specification on the same observations.
Inspect coefficient stability across windows and residual dependence. Quantile regression can examine whether exposures differ in weak-return states, provided the conditional quantile specification and inference are appropriate. An apparent alpha can change when factors, sample period or cost conventions change; it does not by itself identify skill. Report exposures and uncertainty before discussing economic meaning. The workbench’s synthetic financing/confidence data cannot be relabeled as an empirical strategy study.
Chapter basis: Gregoriou & Pascalau (2011), Allen, Singh & Powell, ch. 7, pp. 176–180.
Question: is a change in portfolio risk driven by marginal volatility, dependence or the observation process? For daily analysis, align return intervals and corporate-action adjustments. For intraday analysis, retain timestamps and distinguish transactions from quotes. Compare a simple covariance benchmark against a documented dynamic or noise-aware estimator using the same forecast horizon.
Check matrix symmetry and eigenvalues before interpreting portfolio variance. Track sampling frequency, synchronization rules and missing observations as modeling choices. Evaluate the variance of subsequent portfolio returns as well as statistical covariance errors; a favorable covariance metric need not imply a useful allocation after costs and constraints. A changing correlation estimate is not proof of contagion, and resampling an asynchronous series is not a neutral preprocessing step.
Chapter basis: Terrell & Fomby (2006), Baur pp. 3–10 and Jungbacker & Koopman pp. 186–190; Gregoriou & Pascalau (2011), Mancino & Sanfelici, ch. 1, pp. 3–9.
Question: do spreads or trading activity improve a lower-return-quantile forecast beyond a volatility-only benchmark? Specify the forecast horizon and the information set first. Every predictor must be observable before the forecasted return. Fit on a chronological training window, set the refitting schedule in advance and keep the final assessment period separate from model selection.
Compare exception frequency, exception clustering and quantile loss on the same evaluation observations. A low exception count may reflect an excessively conservative threshold, not superior information. Rare tails yield few exceptions, so failure to reject a test is not proof of calibration. Historical results in the source chapter motivate the research question; they do not establish that liquidity improves every market or period. Keep execution costs and liquidation assumptions separate from a return-quantile forecast.
Chapter basis: Gregoriou & Pascalau (2011), Sanchis-Marco & Rubia, ch. 8, pp. 194–198, 204–208.
N is the number of evaluation observations; rₜ is the realized decimal return; q̂τ,t is its previously issued conditional quantile forecast. ĉτ is a dimensionless exception fraction; its nominal target is τ for continuous returns under the specified conditional quantile. S̄τ is average check loss, in return units, using the loss function defined above. Lower loss favors a forecast only at the same quantile, horizon, units and evaluation sample.
These definitions complement the chapter’s coverage and quantile-regression frameworks. They are not automatically calculated here. Coverage alone does not measure loss beyond the threshold, establish independence or validate causality; use the exception sequence and model assumptions alongside the score.
Method basis: Gregoriou & Pascalau (2011), ch. 7 pp. 179–180, eqs. 7.4–7.7; ch. 8 pp. 204–206. Notation and validation workflows are an original synthesis.
Use a diagnostic to choose the next investigation, rather than treating it as an automatic approval or rejection of a model. The following guide refers to the regression outputs actually available above.
What to inspect: compare the simple and multiple regressions on exactly the same observations. The workbench’s slope-difference identity separates the fitted contribution of z from its sample association with x. A change in slope is evidence of specification sensitivity, not proof that the controlled coefficient is causal.
Next step: justify why the control belongs in the economic model and whether it is known at the decision date. Check that adding it does not create rank problems. If financial factor exposures are the subject, construct the appropriate return/factor dataset rather than using the illustrative investment-index labels.
Factor exposures →What to inspect: a widening residual cloud can motivate a heteroskedasticity review. The auxiliary nR² statistic concerns squared residuals under its own assumptions; neither its magnitude nor the plot supplies an automatic significance conclusion here. Changing to HC1 leaves OLS coefficients unchanged.
Next step: check units, functional form and influential observations before changing the covariance convention. HC1 addresses coefficient covariance under suitable sampling conditions; it does not model a changing conditional variance or correct an endogenous regressor. For ordered financial returns, consider whether a volatility model is the actual research target.
Conditional volatility →What to inspect: retain chronological order and examine the observation frequency, gaps and overlapping return horizons. The displayed Durbin–Watson ratio is descriptive, not a universal serial-correlation test. It is only displayed when time ordering is declared.
Next step: use HAC only when the workbench’s equally spaced, ordered-data requirement and appropriate exogeneity, dependence and moment conditions are defensible. Compare sensible lag choices and disclose them. HAC cannot cure nonstationarity, a wrong mean model or timestamp contamination. If the dependence comes from trade/quote measurement, investigate microstructure before assigning an economic explanation.
Microstructure and sampling →What to inspect: R² and adjusted R² summarize the estimation sample. RMSE here uses n − K in its denominator and retains the outcome’s unit. These are different from a forecast error measured on later observations. A residual plot with no obvious pattern does not demonstrate correct identification or stable parameters.
Next step: define a chronological evaluation window, compare a transparent benchmark and preserve the final holdout for assessment. Select a score matching the target: squared forecast error for a mean, a suitable variance score for volatility or check loss for a quantile. Do not compare differently scaled targets or retrospectively chosen horizons as though they were equivalent.
Study designs and forecast validation →What to inspect: lack of variation and overlapping explanatory information are different from missing or nonnumeric cells. The calculator rejects unusable or near-collinear designs; it does not silently add regularization. Large standard errors can reflect limited information, residual variance or covariance assumptions; they are not evidence of a calculation defect by themselves.
Next step: correct the data issue first and retain the entered observations for review. Use a simpler specification only when economically justified; removing a relevant control solely to obtain a smaller standard error changes the question. For a constant outcome R² is undefined, and a zero standard error does not establish a causal effect. Record sample size and parameter count with the result.
Return to the workbench →State the question and estimand; the dataset, period and units; the specification and covariance convention; the estimated result and uncertainty; the diagnostics reviewed; and what would change the decision. Separate sample evidence, modeling assumptions and unresolved identification questions. If a validation has not been performed, say so explicitly.
Reference basis: Wooldridge (2020), ch. 3, 8 and 12, cited above; Gregoriou & Pascalau (2011), Allen, Singh & Powell, ch. 7 pp. 176–180, and Sanchis-Marco & Rubia, ch. 8 pp. 194–198, 204–208; Terrell & Fomby (2006), Jungbacker & Koopman pp. 186–190. This guide is an original synthesis, not a statistical decision rule.
DECISION FRAMEWORK
A regression coefficient describes a conditional relationship within a specified model. A causal reading needs an identification argument: omitted factors, measurement error and jointly determined variables may connect a regressor to the error term. Adding controls does not automatically remove that connection. Define the population, observation date, variable units and competing explanation before interpreting a scenario as an economic effect. The zero conditional mean assumption concerns the unobserved population error; small fitted residuals do not prove it.
Reference relation and assumptions
Distinguish association from an identified effect; record measurement and sampling conventions; examine alternative specifications and justify exogeneity outside the fitted residuals.
Methodological reference, not a causal estimate generated by this Lab. OLS unbiasedness also requires the other stated sampling, linearity and rank assumptions.
Wooldridge (7th ed., 2020), chapter 3, assumption MLR.4 and equation 3.36, printed pp.82–83 / PDF105–106; chapter 15 §15-1 printed pp.496–498 / PDF519–521.
DECISION FRAMEWORK
Instrumental variables address a regressor correlated with the structural error only when the proposed instrument is relevant and exogenous. Relevance asks whether the instrument explains variation in that regressor; exogeneity asks whether it is uncorrelated with the equation’s unobserved disturbance. A strong first-stage relationship does not establish the second condition. State the exclusion argument, possible direct channels and the population to which the estimate applies. A weak or invalid instrument can make inference less reliable rather than rescue a causal narrative.
Document why an instrument belongs in the first stage but not directly in the outcome equation; inspect first-stage strength and sensitivity to alternative economic channels.
A first-stage test does not certify exclusion or exogeneity. No instrument or estimated causal coefficient is supplied here.
Wooldridge (7th ed., 2020), chapter 15 §15-1, printed pp.496–500 / PDF519–523, specifically equations15.4–15.5 p497 / PDF520.
DECISION FRAMEWORK
Bayesian learning separates the prior belief, the likelihood of a signal and the posterior assessment. Before combining reports, record sample size, signal precision, source independence and contradictory evidence. Several reports repeating one source do not become independent observations. Confirmation bias can affect how information is interpreted; the book's stylized mechanism is a possible explanation, not proof about a particular decision-maker. An investment committee can keep an evidence log and specify in advance which observations would overturn its thesis.
Evidence controls are an AHS application of the learning framework. No objectively calibrated signal quality or experimentally proven debiasing is claimed.
Edward Cartwright, Behavioral Economics, 3rd ed. (2018), §5.1, printed pp.206–207 / PDF237–238; §5.2.1–5.2.3, pp.212–216 / PDF243–247.
RESEARCH NOTE · ORIGINAL SYNTHESIS
Selected readings from the supplied books, connected to reviewable decisions. Historical analysis, model assumptions, author proposals and applicable law are kept distinct.
Educational research, not an entity-specific tax opinion. Local rules, treaties and effective dates require separate specialist review. No client results or tax savings are claimed.
Open seven tax demonstratorsWho owns the activity, signs the contract and earns the revenue?
Oberson separates taxation of enterprises using AI from proposals to give autonomous systems tax capacity. This prevents a policy scenario from being presented as an existing tax status. For an AI deployment, first identify the contracting entity, customer, assets, costs and revenue. A software label does not establish a separate taxpayer or settle the allocation of cross-border profits.
Proposed deliverable: activity and entity map, cost/revenue reconciliation and a list of accounting and tax questions for the relevant jurisdictions. Separate purchased software, services and development work before selecting a treatment.
The 2024 book’s reform arguments are not evidence that AI has tax personality today. No automatic IAS 38 classification or local deduction is asserted.
Xavier Oberson, Taxing Artificial Intelligence, Edward Elgar, 2024, DOI 10.4337/9781035307555.
ch.6, printed pp.58–59 / PDF75–76; ch.10, printed p.187 / PDF204
Published 21 October 2026 · editorial research, no independent validation
Is the supply software, an automated service or a human-led engagement?
Oberson examines AI activities through VAT, especially electronic services and platforms. The useful analytical step is to separate the nature of the supply, the customer’s status and the place of supply. The European Commission likewise explains that the service and customer status must be known before locating taxation. Using the internet or AI is not, on its own, a complete VAT qualification.
Proposed review: compare a self-service API subscription with an expert-led advisory contract. Record supplier, customer status, locations, human intervention and invoices. Produce a localisation and evidence checklist for local tax review.
No VAT rate, exemption, registration duty or permanent-establishment conclusion is automatically determined here.
Xavier Oberson, Taxing Artificial Intelligence, Edward Elgar, 2024, DOI 10.4337/9781035307555.
ch.8, printed pp.105, 146–147 / PDF122, 163–164
Published 21 October 2026 · editorial research, no independent validation
Which base, incidence and behavioural response would a proposal imply?
The author proposes, among other options, an imputed remuneration associated with AI use and discusses the difficulty of measuring it when people and systems work together. This is a policy proposal, not an implemented rate imported into the Lab. A useful comparison tests the tax base, collection mechanism, innovation incentives and distribution of costs among firms, workers and customers.
Research protocol: state a baseline, an explicitly hypothetical tax and alternative pass-through assumptions. Assess administrative cost and possible double counting. Keep budget revenue separate from welfare and employment effects.
No automatic job-loss forecast, empirically estimated incidence or enacted general robot tax is claimed.
Xavier Oberson, Taxing Artificial Intelligence, Edward Elgar, 2024, DOI 10.4337/9781035307555.
ch.9, printed pp.150–151 / PDF167–168; ch.12, printed pp.217–218 / PDF234–235
Published 21 October 2026 · editorial research, no independent validation
Read the mechanism, state its assumptions and connect it to an existing tool. These modules extend the analytical coverage; they do not add calibrated estimators or contractual terms.
Time-series dependence in the conditional mean differs from dependence in volatility. An AR(1) benchmark describes how a rate or return deviation carries into the next observation. Check equal observation spacing, stable coefficients and residual dependence. The long-run expression requires |φ|<1; a plausible fitted line does not establish stationarity. Use chronological assessment and disclose the forecast horizon. No AR estimator is added to the OLS workbench.
y,c,u in outcome units per observation; φ dimensionless; u has zero mean under the model.
Racicot & Théoret, Traité d’économétrie financière (2001), ch.9, pp.230–233.
Explore the related tool or framework →A drifting level can produce misleading regression fit. Differencing changes the estimand from a level to a period-to-period change; detrending removes a specified deterministic component. Neither operation should be chosen only to increase significance. Record how levels are reconstructed, and examine whether the residual dynamics are still misspecified. A stationary-looking plot is not a unit-root test, and no unit-root decision is produced by the current calculator.
Δy retains outcome units per period; level and change forecasts require different interpretation.
Racicot & Théoret, Traité d’économétrie financière (2001), ch.9, ARIMA and trend treatment, pp.245–261.
Explore the related tool or framework →An intercept shift and a change in sensitivity are different hypotheses. A documented break date can be represented by an indicator and interaction, but retrospectively selecting the most favorable date creates selection bias. Compare stable-sample definitions and sufficient observations within each regime. Economic events motivate a specification; the indicator alone does not identify the cause of a change. This is a reading model, not an automated breakpoint search.
D is 0/1; a,d,u outcome units; b,g outcome units per x unit. State the break date.
Racicot & Théoret, Traité d’économétrie financière (2001), ch.4, structural changes, pp.159–160.
Explore the related tool or framework →Resampling approximates uncertainty under assumptions about the data-generating process. Independent observation resampling can destroy the ordering of financial series. Choose a dependence-preserving approach when justified, document the statistic and replication budget, and compare numerical Monte Carlo error with statistical uncertainty. A resampled distribution does not remedy endogenous covariates or a wrong economic model. The book’s examples motivate the method; no bootstrap inference is run here.
B≥2 replications; θ in statistic units; variance in their square. Resampling scheme must be declared.
Racicot & Théoret, Traité d’économétrie financière (2001), ch.6, pp.193–194.
Explore the related tool or framework →Instrument relevance and orthogonality to the unobserved error are separate conditions. GMM chooses parameters that make sample moments small under a stated weighting matrix. More instruments do not automatically improve identification, especially with weak instruments or finite samples. Explain the economic exclusion restriction, rank and covariance estimation. A numerical optimum does not prove that the instruments are valid. This framework is not an implemented IV/GMM estimator.
z instruments; u model error; g moment vector; W compatible positive-definite weighting matrix. Units follow each moment.
Racicot & Théoret, Traité d’économétrie financière (2001), ch.11, pp.326–330.
Explore the related tool or framework →A conditional-variance term in the mean equation expresses a proposed link between risk and expected return. It differs from the GARCH scenario, which only propagates variance using supplied parameters. The sign and strength of the mean coefficient must be estimated under an identified model; historical association is not a universal premium. Define whether risk enters as variance or standard deviation and keep its units consistent. No estimated premium is displayed by the site.
y,u in return units; h conditional return variance; λ inverse-return units in this variance convention.
Racicot & Théoret, Traité d’économétrie financière (2001), ch.10, ARCH-M, p.282.
Explore the related tool or framework →A funded project may be continued to defend a past decision. Selective perception and self-attribution can make successes feel personal and failures external. The book also connects incentives with governance and presents nudging as choice architecture that should preserve freedom and transparency.
Proposed application: independent reviews, prewritten continuation and stopping criteria, comparable option presentation and recorded exceptions. For UI: explain defaults, allow changes and disclose costs alongside risks. Synthesis: make challenge practicable and measure the effects of presentation changes before claiming effectiveness.
Assumptions: comparable options, estimated future flows and explicit risks; test alternative costs, delays and benefits. Positive NPV does not replace capital, liquidity or compliance constraints. Biases are review hypotheses, not evidence of wrongdoing or individual motives.
DECISION FRAMEWORK
Quasi-hyperbolic preferences distinguish the immediate period from all future periods. A lower present-bias parameter can change a decision when an outcome becomes immediate, even if the long-run discount factor is unchanged. This offers a framework for comparing a prior plan with a later choice, not a diagnosis of an investor. Predetermined review dates, explicit decision criteria and documented exceptions are possible governance responses. Personal utility discounting must remain separate from market discount factors used to value bank cash flows.
Reference relation and assumptions
The equation is a linear-utility preference illustration. It does not supply an ALM discount curve, pricing measure or investor diagnosis.
Edward Cartwright, Behavioral Economics, 3rd ed. (2018), §4.2.1–4.2.3, printed pp.172–178 / PDF203–209; reindexed Eq.4.4.