Data preparation
Macro and market series arrive at different frequencies, with gaps, revisions and mismatched calendars. Aligning them by hand is slow, and a quiet mistake here carries through to every result.
An econometrics agent that shows its work.
Describe the research question and bring your data. FinMetric prepares the series, runs the diagnostics, writes the Python or R script, and reports results alongside the output they came from.
Request
Run log
Where research time goes
From specification to script
Turning a model on paper into working code means choosing lags, error distributions and estimators. A small slip can change the estimates without raising any error.
General-purpose chatbots
They can write statistics and citations that look right but were never computed. In research, every number needs an output behind it.
How a run is designed
The language model plans the work and drafts the text. Every number in the report comes from code that actually ran, and you approve the plan before anything is estimated.
-
Agent
Interpret the request
Maps your question to a method, states its assumptions and the checks it will run, then waits for your approval.
-
Computed
Prepare the data
Aligns frequencies, handles gaps and applies transformations in code. Each change is logged.
-
Computed
Check before estimating
Runs the tests the method calls for, such as unit roots, lag selection or ARCH effects, and records the outcome.
-
Computed
Estimate
Fits the model in Python or R and keeps both the full output and the script that produced it.
-
Agent
Report
Drafts the text from the output, with the test or formula named beside each figure. Anything the diagnostics flagged is stated, not smoothed over.
What a run includes
Each method has its own checks before estimation and its own diagnostics after. The log shows which ones ran and which need your attention.
| Method | Before estimation | Reported after |
|---|---|---|
| VAR | Unit root tests, lag-length selection | Stability check, residual autocorrelation, Granger causality, impulse responses |
| GARCH family (GARCH, EGARCH, GJR) | ARCH-LM test on mean-equation residuals, choice of error distribution | Ljung-Box tests on standardized and squared standardized residuals, sign bias test |
| ARDL and NARDL | Order of integration, lag selection, partial-sum decomposition for NARDL | Bounds test, Wald tests for asymmetry, CUSUM stability |
| Panel regression | Cross-sectional dependence test, panel unit root tests | Fixed versus random effects (Hausman), cluster-robust standard errors |
Checked against reference software
The plan is to compare each method's output with reference implementations and published replication datasets before it is released. Differences are investigated, not averaged away.
What stays with you
FinMetric flags problems. Identification, model specification and interpretation remain the researcher's decisions. A run can show that a test failed; it cannot decide what the failure means for your argument.
Where this stands
Status
FinMetric is in early development. The runs on this page illustrate the intended output, and their numbers are placeholders. Request access to be contacted when the beta opens.
Request beta accessWho is building it
FinMetric is developed by Qulub, a finance student who works in R, EViews and Python on quantitative finance and econometrics research. He has also worked in equity research at a securities firm.
qulub@finmetric.xyz