FinGraph — the complete researcher's manual

FinGraph turns one table of firm financial data into networks, clusters, machine-learning predictions, and a fully tested research model — free, and entirely in your browser. This guide explains every control, shows how to read every output, gives worked examples, and provides paste-ready sentences for writing up your results.

Open FinGraph → ⬇ Download PDF

Diagrams below are schematic illustrations of each panel. Nothing you upload leaves your computer — every calculation runs locally in your browser, so your data stays private.

What FinGraph does

FinGraph is built for panel data: each row is a firm observed in a year (a “firm-year”), and each column is a financial variable — a balance-sheet line item, an income-statement figure, or a ratio. From that single table it offers five families of analysis, all sharing the same data and settings.

The five families are: (1) association networks that map how variables co-move; (2) ratio networks over auto-derived financial ratios; (3) clustering with per-cluster finance analysis; (4) machine-learning prediction of a performance target; and (5) a research-model builder for hypothesis testing with regression, moderation and mediation.

Who it is for

FinGraph suits accounting, finance and management researchers, graduate students, and instructors who work with company financial panels (CSMAR, Compustat, Orbis, hand-collected data, or teaching datasets). No coding or statistics software is required, but the outputs are rigorous enough to report in a paper: significance tests, standardised coefficients, VIF, out-of-sample validation, bootstrap confidence intervals, and reproducible method statements.

Quick start (3 steps)

1. Load your data. On the tool page, drag a CSV or Excel file onto the drop zone, or click ✨ Try a sample to explore with demo data first.
2. Check the columns. In Configure, confirm FinGraph identified the right Firm ID, Firm name and Period/Year columns. Every other numeric column becomes an analysis variable.
3. Run an analysis. Build a network, or scroll to Predict firm performance or Research model builder and press the run button. Every result can be interpreted, expanded and exported.

Choosing an analysis

Match your question to the right tool:

Preparing your data

FinGraph expects a “long” panel — one row per firm-year:

File requirements

• Format: CSV or Excel (.csv, .xlsx, .xls). Large Excel files stream in the browser; for very large exports (hundreds of MB) a CSV subset (one year, or selected firms/variables) loads fastest.
• Identifier column: a firm code (e.g. stock code, GVKEY, ISIN). FinGraph uses it to group firm-years and to split by firm.
• Period column: a year or date. Needed for within-firm and growth transforms and for the timeline.
• Total Assets: include it if you want common-size transforms and automatic ratio derivation. FinGraph recognises the CSMAR code A001000000 or any column labelled “total assets”.
• Works with the major providers out of the box. FinGraph auto-recognises the column conventions of Compustat/WRDS (mnemonics like at, lt, ceq, sale, ni; identifiers gvkey/datadate), CSMAR (A00… codes; Stkcd/Accper), Worldscope (WC… codes), and CRSP (permno). Orbis, Datastream, Capital IQ and EDGAR exports match by their readable item names. So identifiers, periods, labels, ratios and distress scores are derived automatically — no manual mapping. See the “Works with the major datasets” panel on the load screen for per-provider export tips, and correct anything in the Configure step.

The Configure step

After loading, FinGraph auto-detects the ID / name / period columns. Verify them — if a column is mis-detected, pick the right one from the dropdowns. In Variables in the network you can keep All numeric columns or the Top 30 by coverage (coverage = how many rows have a value).

Reshape / query your data with SQL

The load screen has an optional 🛢️ Reshape / query your data with SQL panel (it appears after a file loads). Use it to select, filter, rename, aggregate or pivot your table into exactly the firm-year rows you want — everything runs in your browser; nothing is uploaded.

• The loaded table is referenced as ?. Wrap awkward column names in [brackets], e.g. [A001000000].
• A helper function num(x) coerces messy text to a number (handy inside SUM/AVG).
• Three preset buttons cover the common cases: keep recent years, select & rename columns, and aggregate by firm-year.

Standardize variables (at the start)

The Configure step has a Standardize variables control that rescales every numeric variable once, up front, and feeds the standardized values into every analysis (network, heatmap, dashboard, prediction, regression, causal designs). Choose the method your field prefers:

Selecting specific years

Below standardization, the Years in the analysis list has a checkbox per year in your data. Tick the years you want and the whole tool — the data-overview dashboard, network, heatmap, predictions, regression and causal designs — is restricted to those firm-years. Use All to reset or Latest to jump to the most recent year. This is the global filter; the network card also has a year slider that animates through the selected years one at a time. Standardization is computed within the selected years, so “z-score” means relative to the sample you kept.

Transforms explained

Raw balance-sheet levels are dominated by firm size, so big firms have big everything and every variable looks correlated (a “size hairball”). A transform removes that. Choose based on your question:

Missing values & outliers

• Missing data: analyses use listwise-complete observations — a firm-year is dropped from a model if the outcome or any included variable is missing. The reported N is the complete-case count. Networks use pairwise-complete correlations with a minimum-pairs guard so sparse cells don’t create spurious edges.
• Outliers: extreme ratios (from tiny denominators) can distort correlations and regressions. Consider winsorising or trimming extreme firm-years in your source file before uploading; FinGraph flags instability indirectly through wide standard errors and high VIF.
• Coverage: if a variable is mostly empty it will be dropped or weaken a model. Use “Top 30 by coverage” to prioritise well-populated variables.

Data overview dashboard

Between Configure and the network, the Data overview card gives you an at-a-glance exploratory picture of your panel — the kind of descriptive dashboard you would otherwise build by hand — so you can sanity-check the sample and see how each variable behaves over time before modelling.

What it shows

• Sample counts (KPI strip): number of firms, years, firm-year observations, groups (e.g. industries), and variables. These are the counts you report in a “sample” paragraph.
• Sample composition (bar chart): firm-years per industry, firm-years per year, or firms per industry — switch with Counts bar chart.
• Featured variable (enlarged): a large line chart of one “main” variable over time, one line per industry. Pick it in Featured variable.
• Small-multiples grid: every selected variable as its own mini timeline, laid out in a grid (set the number of panels and columns). Lines are coloured by industry using a single shared legend, so colours mean the same thing in every panel. Hover a panel for a quick ⬇ PNG, or click it (or the ⤢ button) to open it full-size and interactive — with a legend and hover tooltips — and download a 300-DPI PNG from there. Every expanded chart in FinGraph now has a PNG button.

Controls

1 · Association network

What it answers: which financial variables move together across firms, and how they group. Strongly correlated variables sit closer, are joined by thicker links, and cluster by colour; negative correlations are dashed red.

Steps

1. Keep the Variable network tab selected.
2. Choose a Transform (common-size is a good default) and a Correlation method (Pearson for linear, Spearman for rank/robust).
3. Set Minimum |correlation| and Max variables, then press ⚙ Build network.
4. Open Advanced — filtration & model to choose how edges are kept and whether to use partial correlations (see the control table below).
5. Tune Community resolution and Colour nodes by; toggle hulls, rings and density for readability.
6. Press 🏷️ Name clusters (AI) and 🧠 Interpret network (AI) for labels and a written read-out.

Every control explained

Reading the output

The stats bar reports nodes, edges, density, number of clusters, and modularity Q (clustering quality — higher is cleaner). The insight panel lists the strongest correlations, cluster membership, and pivotal variables (high betweenness — they bridge clusters, marked ◆). A variable that bridges two clusters is often an integrator (e.g. total current assets linking working-capital items).

1b · Correlation heatmap & lower-triangle matrix

What it answers: the same correlations as the network, shown as a colour matrix — the familiar correlation table you put in a paper. It appears in the Correlation heatmap card below the network and redraws whenever you build a variable or ratio network. Blue = positive, red = negative, white ≈ 0.

Exports: a 300-DPI PNG, a lower-triangle CSV formatted as an APA correlation table (numbered variables, blank upper half, 1s on the diagonal, stars included), and a full-matrix CSV.

2 · Ratio network

What it answers: how the standard financial ratios relate to one another — the way analysts actually think. FinGraph auto-derives up to 20 ratios from your columns and networks those.

* Turnover and profitability ratios appear only when income-statement items (revenue, net income) are in your file — the ratio set adapts to your data.

1. Click the Ratio network tab.
2. Press ⚙ Build network. Ratios that co-move cluster into families (liquidity, leverage, profitability…).
3. Use the same clusters, exports and AI interpretation as the variable network.

3 · Clusters & per-cluster finance analysis

What it answers: what each cluster means in finance terms. After any variable or ratio network, the Cluster finance analysis panel turns each cluster into a finance object.

1. Build a network, then scroll to Cluster finance analysis and pick a cluster from the dropdown.
2. Read the member table: each item’s share of total assets, mean and SD, and its PC1 loading (how strongly it defines the cluster).
3. Note the cluster’s cohesion (average internal |r|), its total weight (% of assets), its hub (most connected member), and the composite-index equation — the first principal component, a single index you can compute per firm-year to summarise the cluster.
4. Press 🧠 Interpret this cluster (AI) for the economic meaning, the ratios it maps to, and a testable hypothesis.

Reading the output

• Cohesion ≈ 0.9 means the members move almost in lockstep — a tight construct.
• Cluster weight tells you how much of the balance sheet the cluster occupies (e.g. a working-capital cluster at 40% of assets).
• PC1 “explains X% of variance”: a high number (say 80–95%) means one underlying factor drives the whole cluster, so the composite index is a faithful summary.

4 · Firm (peer) network

What it answers: which firms are structurally similar. Each firm is reduced to its average standardised financial profile; firms with similar profiles are linked and cluster into peer groups.

1. Click the Firm network tab and press ⚙ Build network.
2. Read the peer-group clusters; the cluster panel shows what makes each group distinctive (variables higher and lower than the overall average).

Comparing peer groups & analysing one cluster

After building the firm network (and optionally naming the clusters with AI), the cluster panel gives you two extra tools:

1. 📊 Compare peer groups (ANOVA) — runs a one-way ANOVA across the clusters for every variable and reports, for each, the F-statistic, p-value, η² (the share of variance explained by cluster membership — an effect size), and the per-cluster means. It tells you exactly how the peer groups differ statistically and which variables separate them most.
2. → Use this peer group in the research model — sets the model's firm sample to the selected cluster's firms, so every analysis lower down the page (research model, moderation/mediation, prediction, and the causal designs) runs on just that peer group. Combine it with the same tool on another cluster to compare results across groups.

5 · Predict firm performance (machine learning)

What it answers: can a performance measure be predicted from the other variables, how accurately, and what drives it? Results report out-of-sample accuracy on a held-out test set — the honest test of a predictive model.

Steps

1. Choose a Target — what to predict. Options include derived growth measures (asset/revenue/equity YoY), any ratio (ROA, ROE…), or any raw item.
2. Pick the Task: Regression (predict the value) or Classification (high vs low performer, split at the median).
3. Choose the Predictors (ratios, raw items, or both) and the Model (see the model table below).
4. Set the Horizon: Same period (explain) or Next period t+1 (forecast — year-t features predict the target at t+1).
5. Keep Split by firm ticked (holds out whole firms so the test can’t “cheat” by seeing the same firm in training and test), set the test %, then press ⚙ Train & evaluate.

Which model?

Reading the output

• Regression: Test R² (share of out-of-sample variance explained), Adjusted R², Test RMSE (typical error), and Train R² (in-sample). The actual-vs-predicted plot should hug the diagonal.
• Classification: Accuracy, ROC AUC (0.5 = chance, 1.0 = perfect), Precision, Recall, and the ROC curve.
• Importance: permutation importance (for the NN) or standardised coefficients (for linear/ridge) rank what matters, with direction.

6 · Research model (multiple regression)

What it answers: does variable X relate to an outcome, controlling for other variables? This is the core hypothesis-testing tool: specify a model, test it, and get publication-ready statistics.

Steps

1. Pick the Outcome (dependent variable) — e.g. ROA, or asset growth.
2. Optionally restrict the sample: open Firms — all and tick the firms to include (filter / All / None), or — after building the Firm network — pick a whole peer-group cluster from the dropdown to select all its firms at once. The header shows “N of M”. Use this for subsamples (an industry, a size band, a peer group).
3. Tick your Predictors and any Control variables. Controls are entered but reported separately, so your hypothesis variables stay in focus.
4. Choose the Horizon (same period, or next-period t+1 for a predictive test).
5. Press ⚙ Test model.
6. Read the diagram, statistics table, fit cards and hypothesis tests; press 🧠 Interpret (AI) or ⬇ Download full report.

Reading the statistics table

Reading the fit & hypotheses

R² / Adjusted R² is how much of the outcome the model explains; the F-test asks whether the model as a whole beats no predictors (a tiny p means yes). Each hypothesis line states whether a predictor is supported (significant, with its direction) or not supported.

6b · Panel methods, robust inference & scores from the literature

Because financial data are firm-year panels, top accounting and finance journals rarely use plain OLS. The research-model card now offers the field-standard toolkit, added after a review of methods used in The Accounting Review, Journal of Accounting Research, Journal of Finance and the Review of Financial Studies.

Estimator — fixed effects

The Estimator selector controls unobserved heterogeneity via the within transformation:

Standard errors

OLS standard errors are biased in panels because residuals correlate within firm (Petersen 2009). The Standard errors selector offers Classical, Robust (HC1) for heteroskedasticity, and Clustered by firm — the expected default in accounting and corporate finance. The estimator line under the fit cards records your choice and the number of clusters.

Winsorise

The Winsorise selector caps each variable's tails at the 1st/99th (or 2.5/97.5, 5/95) percentiles before estimation — the near-universal fix for the heavy tails of financial ratios.

Descriptives (Table 1 & Table 2)

The 📋 Descriptives button produces the two tables every empirical paper opens with: Table 1 (N, mean, SD, min, P25, median, P75, max for the outcome and predictors) and Table 2 (a starred correlation matrix). Both are included in the downloadable report.

Distress, quality & manipulation scores

When your data include the required income-statement / cash-flow items, FinGraph auto-derives the classic score models as variables you can use as an outcome, predictor or control:

If the inputs are missing, the score simply does not appear — the set adapts to your data. Scores that need a prior year use the panel's lag automatically.

Difference-in-differences (event study)

The Difference-in-differences panel estimates the causal-style effect of an event (a policy, reform or shock) on a treated group versus a control group. Pick a treatment-group indicator (a 0/1 column marking treated firms) and an event year; FinGraph forms Treated, Post and Treated × Post and estimates Y = β₀ + β₁·Treated + β₂·Post + β₃·(Treated×Post) + controls with firm-clustered SEs. The DiD estimate is β₃. You also get a 2×2 means table (treated/control × pre/post) whose difference-of-differences equals β₃, and a treated-vs-control trend chart with the event marked — use it to check the crucial parallel-trends assumption (the two lines should track each other before the event).

Causal designs — IV, PSM, RD, event study

The Causal designs panel adds four more identification strategies. Each needs a design input you supply; all report an estimate with inference and (where useful) a diagnostic chart.

7 · Auto-find predictors

What it answers: out of many candidate variables, which significantly relate to the outcome? 🔎 Auto-find predictors enters all available predictors and reports which are significant.

1. Set the Outcome (and any controls / firm selection). You do not need to tick predictors — Auto-find uses them all.
2. Press 🔎 Auto-find predictors. FinGraph then:
   • fits the full model (every predictor together) and makes it the active model — significant predictors appear solid in the diagram, non-significant grey dashed;
   • runs a univariate screen — each predictor tested one at a time with the controls — and reports its p-value alone versus in the full model;
   • searches predictor combinations (best-subsets when feasible, else forward stepwise) and recommends a parsimonious model, ranked by adjusted R² and BIC.
3. To adopt a recommendation, tick those predictors and press ⚙ Test model.

p “alone” vs p “full” — why both?

A predictor can be significant alone but not in the full model (its effect was shared with correlated predictors), or vice-versa. Comparing the two columns reveals which relationships are robust to controlling for everything else. Predictors significant in both are the strongest candidates.

8 · Moderation (interaction)

What it answers: does the effect of X on Y depend on a third variable W? A significant X × W interaction means yes.

1. Set the Outcome (Y) at the top of the card.
2. Open Moderation / Mediation, choose Moderation, pick X (predictor) and W (moderator), and press ▶ Run test.
3. Read the interaction row: a significant X × W term (p < .05) means W moderates the X → Y effect.
4. Read the simple slopes and the plot: the X → Y line is drawn at low (−1 SD), mean, and high (+1 SD) W. Diverging lines = moderation.

9 · Mediation

What it answers: does X affect Y through a mechanism M? Mediation decomposes the total effect into a direct path and an indirect path X→M→Y.

1. Set the Outcome (Y). Open Moderation / Mediation and switch the test to Mediation — the second box becomes the mediator M.
2. Pick X and M and press ▶ Run test.
3. Read the paths: a (X→M), b (M→Y controlling X), c (total X→Y) and c′ (direct). The indirect effect is a×b.
4. Check the tests: the Sobel z/p and — more importantly — the bootstrap 95% CI (800 resamples). If the CI excludes 0, the indirect effect is significant.
5. Read the verdict: full mediation (indirect significant, direct not), partial (both significant), or none.

10 · Exports & the full report

• Expand: every chart has an ⤢ Expand control for a full-screen view (Esc or click outside to close).
• Network exports: PNG (300 DPI), SVG, edge list (Gephi/VOSviewer), correlation matrix, node-cluster CSV, GraphML, Pajek (.net), and a paste-ready Methods + settings paragraph.
• Full report: ⬇ Download full report compiles everything you have run — the research model (diagram + tables), moderation/mediation, predictive results (with charts), and the network — into one self-contained HTML file. Open it in any browser and print to PDF for an appendix.

Interpreting significance & effect sizes

Two questions matter for every result: is it significant? (unlikely to be chance) and is it large? (does it matter).

• p-value: the probability of seeing an effect this large if the true effect were zero. p < .05 is the common threshold, but with large samples even trivial effects become “significant” — always look at size too.
• Standardised β / correlation r: ≈ .10 small, .30 medium, .50 large (Cohen’s rough guide).
• R²: the share of variance explained. In firm-level social-science data, R² of .10–.30 is common and respectable; very high R² often signals a mechanical or circular relationship.
• Confidence intervals convey both: a narrow CI far from 0 is a precise, meaningful effect; a wide CI spanning 0 is inconclusive.

Statistics glossary

Reporting templates

Copy, then fill in your numbers (or use the tool’s Methods + settings and Download full report):

An example end-to-end workflow

1. Load a CSMAR-style balance-sheet export (or the built-in sample).
2. Explore structure: build a common-size variable network, press Name clusters (AI), and read the per-cluster finance analysis.
3. Ratios: switch to the ratio network to see liquidity / leverage / profitability families.
4. Form a question: e.g. “Does leverage predict next-year asset growth, controlling for size?”
5. Test it: in the research model builder, set outcome = Asset growth (YoY), predictor = Debt-to-assets, control = Total Assets, horizon = next period; press Test model.
6. Explore drivers: run Auto-find predictors to see all significant correlates; check p-alone vs p-full and the parsimonious model.
7. Mechanism & boundary: test mediation (does liquidity carry the effect?) and moderation (does size change it?).
8. Predict: train a neural network to forecast next-year growth and compare its out-of-sample accuracy to the regression.
9. Report: press Download full report, print to PDF, and paste the method statement into your paper.

Troubleshooting & FAQ

Good practice

• Transform first. Always remove the firm-size effect before interpreting correlations or coefficients.
• Association ≠ causation. Regression, moderation and mediation on observational panels describe associations, not proven mechanisms. Use lags (t+1), theory, and — where possible — natural experiments to support causal claims.
• Beware data dredging. Searching many predictors and models inflates false positives; confirm exploratory findings on theory or new data.
• Mind multicollinearity. Watch VIF; two near-identical predictors make individual coefficients unstable.
• Report your settings. Use Methods + settings and the full report so your analysis is reproducible.
• Not investment advice. FinGraph is a research and teaching tool.

Open FinGraph and try it →