1 Load firm financial data

CSV or Excel. Rows = firm-year observations; columns = financial variables. Auto-recognises the major providers — Compustat/WRDS (at, lt, ceq…), CSMAR (A001101000…), Worldscope (WC02999…) and readable item names — so ratios and distress scores are derived automatically.

🕸️

Drop a file here or click to browse

CSV · XLSX · XLS

Very large exports (100s of MB) may not fit in a browser — export a subset first (one year, or selected firms/variables) or a CSV.

📚 Works with the major finance & accounting datasets — how to export

FinGraph auto-recognises the column conventions of the main providers, so exports load with labels, ratios and distress scores derived automatically. Export one row per firm-year (long panel) as CSV or Excel:

Dataset	ID · period columns	How to export
Compustat / WRDS (Fundamentals Annual)	gvkey · datadate/fyear	WRDS → Compustat → Fundamentals Annual → download CSV (mnemonics at, lt, ceq, sale, ni… are auto-labelled)
CSMAR (China)	Stkcd · Accper	FS_Combas / income / cash-flow query → Excel; A00… codes auto-resolve
Refinitiv Worldscope / Datastream	code · year	WC-codes (WC02999…) are recognised; Datastream readable names are matched by label
Orbis / BvD, Capital IQ	company id · year	Export with English item names (Total assets, Total liabilities…) — matched automatically
SEC EDGAR (Financial Statement Data Sets)	cik · period	Pivot the tag/value files to one row per firm-year before uploading
CRSP (returns/prices)	permno · date	Best merged with Compustat (CCM); FinGraph reads the panel

Any tidy panel works even if your provider isn't listed — FinGraph matches readable column names (e.g. “Total Assets”, “Net Income”). You can also correct any column in the Configure step. Datasets are your own licensed extracts; FinGraph never uploads them.

Multi-row headers are handled automatically: CSMAR/Datastream exports stack a field-code row, a description row and a “unit” row — FinGraph keeps row 1 as the header and drops the description/unit rows so only real firm-year records load.

2 Configure

Firm ID Firm name Period / year Standardize variables — applied to all analyses

Scales every numeric variable before analysis. Z-score / min–max don't change correlations (they're scale-invariant) but matter for machine learning, firm similarity and the timelines; robust / log / rank also reshape distributions. Leave on None for distress scores and common-size ratios, which expect raw values.

Years in the analysis — applies to all charts & models

Tick the years to include. The selection filters the dashboard, network, heatmap, predictions and models to those firm-years.

Industry / group in the analysis — applies to everything

Pick the grouping column (industry, sector…) and tick which groups to include. Everything below is restricted to the selected firms.

Variables in the network

Data overview — timelines & sample counts

A dashboard of your panel: how many firms, years and firm-year observations you have, and how each variable moves over time. Lines are coloured by industry (or any grouping you pick). The featured variable is enlarged; the rest appear as a small-multiples grid below.

Colour / group by — e.g. industry Featured (enlarged) variable Line = firms' Counts bar chart Panels in grid 24 Grid columns 6 Standardise each panel (z-score) — compare shapes Show mini-panel titles

Grouping options are the categorical columns in your file (industry, sector, exchange…). Click any mini-panel to enlarge it.

Sample composition

Featured variable over time

Every variable over time — small multiples

Hover any mini-panel for its own ⬇ download.

3 Network

Nodes are variables. An edge means the two variables are correlated across firms; the stronger the correlation, the closer and thicker the link. Clusters of co-moving variables are coloured.

Transform — removes firm-size effect Size control — partial correlation Correlation method Minimum |correlation| 0.50 Max variables 40 Edge sign

Advanced — filtration & model

Edge selection — how edges are kept Partial-correlation network (GGM) — conditional dependence Community resolution 1.00 Min. pairs per edge 8

Node size by Colour clusters (communities) Show labels Cluster hulls Cluster rings + names Density glow Cluster separation 0.0 Colour nodes by

Correlation heatmap — matrix view of the network

The same correlations shown as a colour matrix. Blue = positive, red = negative, white ≈ 0. Use the lower-triangle view for a clean publication figure. Built from the variables/ratios in the network above — rebuild the network to change them.

Show Diagonal Order variables Matrix Show values in cells Significance stars (* .05 ** .01 *** .001) Use short codes for labels

Tip: the lower-triangle + values view is the classic correlation table for papers.

Predict firm performance — machine learning & predictive analytics

Choose a performance target (e.g. ROA, ROE, asset growth) and let FinGraph learn to predict it from the other financial items or ratios — with a linear model, ridge regression, or a neural network. Everything trains in your browser; your data never leaves it. Results report out-of-sample accuracy on a held-out test set.

Target — what to predict Task Predictors (features) Model Horizon Test set held out 30% Split by firm — avoids leakage

Research model builder — build a model & test it on your data

Specify a research model — an outcome (e.g. firm performance) explained by predictors and control variables — and test it with multiple regression. You get the model diagram, a full statistics table (B, SE, β, t, p, VIF, 95% CI), model fit (R², adjusted R², F), and hypothesis support. Download the whole thing as a report.

Outcome (dependent variable) Horizon Estimator — panel method Standard errors Winsorise — tail outliers

Firms — all

Predictors (independent variables)

Control variables — optional

Moderation / Mediation

Test X — predictor W — moderator

Outcome Y = the dependent variable above; ticked controls are included.

Difference-in-differences

Treatment group — a 0/1 column Event year — treatment starts

Outcome Y = the DV above; SEs clustered by firm. Check the pre-event trends look parallel.

Causal designs — IV · PSM · RD · event study

Design

Endogenous XInstrument Z

Outcome Y = the DV above; controls = ticked control variables.

Earnings quality & forensics — discretionary accruals & Benford's law

Two accounting-research tools. Discretionary accruals (Jones / modified-Jones) estimate the "abnormal" part of accruals — a widely used earnings-management proxy. Benford's law screens the leading digits of reported figures for anomalies. Needs income & cash-flow columns (Net income, Operating cash flow, Revenue, Receivables, PPE).

Accruals (earnings management) Model Total accruals Estimate by Min firms per group 8

Benford's law — column

Discretionary accruals |DA| = the size of earnings management; higher = more aggressive. Benford flags digit patterns unlikely in unmanipulated data.

Dimensions & clusters — PCA / factor analysis & firm clustering

Reduce your variables to the underlying factors with principal component analysis (scree, loadings, variance explained), and group firms by their financial profiles with k-means clustering (silhouette + a principal-component map). Uses the variables selected in Configure.

Principal component analysis

Firm clustering — number of clusters (k) 3

PCA finds latent factors driving the balance sheet; clustering finds peer groups. The scatter maps everything onto the first two components.