How to use StatEVN
1. Getting started
StatEVN turns a table of data into a statistical model and explains the result in plain language. You do not need to write code.
- Upload a CSV, Excel or JSON file.
- Say what you want to learn, for example "what drives revenue" or "which customers are likely to churn".
- Pick the outcome (the thing you want to explain or predict) and the factors you think matter.
- Run the analysis. StatEVN cleans the data, chooses a model, fits it and shows the results, charts and warnings.
The whole process usually takes under a minute for a normal file.
2. Preparing your file
The best file is one table with a header row, one row per case (a customer, a day, an ad group) and one column per variable.
- Formats: .csv, .txt, .xlsx, .xls and .json.
- Size: the number of rows allowed per file depends on your plan (see Plans and limits).
- Pivot-table exports: if your file has months or categories written across the top (a two-row header), StatEVN can detect that and turn it into a flat table with one row per item and period. This is available on the Starter, Analyst and Pro plans. Ordinary files are never changed by this step.
- Numbers: avoid putting units or symbols inside numbers where you can, for example write 1250 rather than "1,250 USD". StatEVN tries to read such values, but a clean column gives the best results.
3. Data cleaning
Real data is rarely ready for a model. Before anything is fitted, StatEVN runs an automatic cleaning step and then shows you exactly what it did in the Data cleaning summary on the results page. Nothing happens silently: every change is listed there.
| What StatEVN finds | What it does |
|---|---|
| Numbers stored as text (for example "1250" or "12 %") | Converts the column to numbers. If only a few values (up to 5%) cannot be read, they are treated as missing. If 5% to 50% cannot be read, the column is still converted and you get a warning. If more than half cannot be read, the column is kept as text and treated as categories. |
| Dates written in different ways (for example 06/08/2018 06:00 in one row and 09-09-2018 2:44 AM in another) | Every date in the column is rewritten in one format: dd/mm/yyyy hh:mm:ss (dd/mm/yyyy if no row has a time). Fractions of a second are rounded to whole seconds. StatEVN also tells you whether the rows run from oldest to newest, newest to oldest, or are not in order. Dates such as 06/08/2018 can be read as day/month or month/day: StatEVN reads them as day/month unless other rows prove otherwise (for example 09/25/2018), and warns you when it had to assume. Values that cannot be read as dates are left as they were and counted. |
| Amounts with currency signs (for example €1,200 or (300) for a negative) | If the column uses one currency, the sign is removed and the column becomes numeric. If it mixes currencies (for example EUR, USD and GBP), StatEVN leaves it unchanged and warns you, because amounts in different currencies cannot be added or averaged. Convert them to one currency first. |
| A column that mixes numbers and words | You get a warning with examples of the text found, so you can decide whether the column should be all numbers or all categories. |
| Infinite values (for example a ratio divided by zero) | Treated as missing values, because they break every model. The affected columns are listed. |
| Duplicate rows | Removed only when that is safe. With few columns, identical rows are often a coincidence rather than a mistake, so StatEVN checks the chance of that first. If more than half the rows look like duplicates it removes nothing and asks you to check the data. The decision and the reason are shown. |
| Columns that are mostly empty | A column with half or more of its values missing is dropped, because filling it in would mean inventing most of it. |
| Missing numbers | It depends on whether your rows follow time. Rows in date order (one row per day, month or year): the gap is filled from the neighbouring rows with a weighted moving average, where closer rows count more and the memory length adapts to how jumpy the column is. Any other order (a list of customers, products or countries): the row above says nothing about the row below, so the gap is filled with the column's median, its middle value. The number of filled values per column is shown. |
| Missing yes/no values and ratings (for example 1 to 5) | Each row is placed in the group whose other values look most like that row. The other columns are first put on the same scale, so a large column such as revenue cannot outweigh a small one such as a 1 to 10 score. If a row has half or fewer of its other values filled in, there is too little to go on and the row is removed. The count of removed rows and the reasons are listed. |
| Missing text values | Filled with the most common value in that column. |
| Outliers | Flagged, never deleted. A value far outside the usual range is counted and reported, because only you know whether it is an error or a real, important case. |
| Text categories (region, plan, channel) | Turned into 0/1 indicator columns, one per category. One category is left out as the reference, and effects are described relative to it. |
| Text columns with very many different values (names, IDs) | These cannot be used as categories, so they are left out of the model. For long lists such as ad-group names StatEVN offers grouping (see below) so the information is not lost. |
| Shares that always add up to 100% (or 1) | Such columns are perfectly linked to each other, which makes regression results meaningless. StatEVN leaves out the largest share as the reference and tells you. |
Grouping long text columns
If a column has hundreds or thousands of different values, StatEVN suggests a way to turn it into a handful of groups:
- By name pattern groups names that share structure, such as the same leading word, with ids removed and typos merged.
- By behaviour groups values by how they perform in your numeric columns.
- Top values keeps the most frequent values and puts everything else under "Other".
- None leaves the column out of the models.
You see a preview of each option and choose before the analysis runs.
What cleaning cannot know
Cleaning fixes the technical problems it can detect. It cannot know that a revenue figure is in the wrong currency, that a date is wrong, or that an outlier is a typo. Check those at the source. The outlier and warning lists in the cleaning summary are the best place to start.
Where to see it
Open the Data cleaning summary block on the results page. It lists the original and final size, duplicates, dropped columns, filled values per column and the method used, removed rows, outliers, and converted columns.
4. Choosing the outcome and the factors
- Outcome: the column you want to explain or predict, such as revenue, sales or churn.
- Factors: the columns you think influence it. Include what you believe matters, but avoid columns that are just another form of the outcome (for example a total when the outcome is one of its parts). The reliability check will warn you if it sees this.
- Your goal in words: a short sentence helps StatEVN choose the right kind of model, for example "predict which customers will leave" points to a yes/no model.
- Comparing groups: to test whether a number, such as ARPU, differs between groups, such as regions, put the number in Outcome and one text column in Factors, or say "compare" or "ANOVA" in your goal. StatEVN runs a t-test for two groups or ANOVA for three or more, shows which pairs differ, and tells you how reliable the result is. To compare several numbers at once, put the other numbers in Factors next to the group column: you get one summary table with a test per number, adjusted for testing several, and you can click a row for its full comparison.
- No outcome at all: if you only want to find groups of similar rows, choose grouping or clustering instead.
5. Amateur and Advanced mode
Amateur mode gives you the answer in plain language: what matters, in which direction, how reliable it is and what to watch out for.
Advanced mode (Analyst and Pro plans) adds the technical detail an analyst expects: full statistics tables, confidence intervals, and diagnostic tests, for example for uneven variance and for factors that move together.
6. How StatEVN chooses a model
StatEVN looks at your outcome, your goal and your data, picks a suitable model, and tests several others on the same data so you can see whether the choice holds up (the model comparison table).
| Model | Used for |
|---|---|
| Linear regression (OLS) | A numeric outcome such as revenue. Easiest to read: every factor gets an effect per unit. |
| Poisson, Binomial, Gamma, Gaussian (GLM) | Counts, proportions, positive skewed amounts and similar outcomes that a straight line fits badly. |
| Logistic regression | A yes/no outcome such as churn. Effects are described as changes in the odds. |
| Decision tree | Simple if-then rules. Easy to follow, handles non-linear patterns. |
| Gradient boosting and XGBoost | Often the most accurate for complex patterns, but harder to explain. You get factor importance and an effect curve instead of coefficients. |
| t-test and ANOVA (compare groups) | Checking whether the average of a number, such as ARPU, really differs between groups, such as regions, and which groups differ. |
| A/B test (two-proportion test, Welch t-test, sample size) | Checking whether a version really beat the control in a test, with confidence intervals and lift, and planning how long a test must run. Starter, Analyst and Pro plans. |
| K-Means, hierarchical, DBSCAN clustering | Finding groups of similar rows when there is no outcome to predict. |
| ARIMA | Forecasting a series over time. |
On the Free plans the models are limited to OLS, Gradient Boosting Regressor, Decision Tree, Logistic Regression, K-Means, group comparison (t-test and ANOVA) and ARIMA. The Starter plan adds the GLMs (for counts, shares and amounts that cannot go below zero) and Exponential Smoothing. If your data needs a model outside your plan, StatEVN uses the closest available one and tells you.
7. Reading your results
- R² (test R²) says how much of the variation in the outcome the model explains on data it has not seen. It is held out from the 20% of rows the model never learned from. 0 means nothing, 1 means everything. Be suspicious of values very close to 1.
- "Significant" and the p-value tell you whether an effect is distinguishable from chance. A small p-value (below 0.05) means the effect is unlikely to be luck. A factor that is not significant adds little and may be removed. StatEVN suggests this and lets you re-run without it.
- Effect sentences such as "when posts increases by 1 unit, revenue increases by about 137" describe the effect with the other factors held the same. If several factors change together in real life, the combined effect can differ.
- Logistic results describe changes in the odds of the outcome, not a plain percentage chance.
- Correlations show how each variable relates to the outcome on its own. The model looks at all factors together, so the two can disagree, and that is normal.
- Model comparison shows how other models did on the same hold-out data. If a simpler model performs about the same, StatEVN prefers it because it is easier to explain.
Statistical results describe patterns in your data. They do not prove cause and effect.
8. Charts
The Charts section (Starter, Analyst and Pro plans) draws the model's performance on the rows it did not learn from. The Starter plan shows the result charts; the charts that check the model (errors and the ROC curve) are on the Analyst and Pro plans.
Compare models (Pro plan): after you choose your outcome and factors, click Compare several models instead. Pick at least two kinds of model. StatEVN trains them all on the same rows, scores them on the same rows none of them has seen, and shows which predicts best, which is easiest to explain, and whether they agree on which factors matter. It counts as one analysis. It works when you predict a number or a yes/no outcome. When you are grouping items into clusters, you compare K-Means, Hierarchical and DBSCAN instead, and also see whether the methods agree on which items belong together.
Choosing a model yourself (Analyst and Pro plans): on the variables screen, Model choice lets you pick Random Forest, Gradient Boosting, XGBoost or Decision Tree for a prediction goal, and Exponential Smoothing or Prophet for a forecast. Leave it on Automatic and StatEVN picks the best fit. A Random Forest lets many decision trees vote, which makes it steady and hard to over-tune. Exponential Smoothing gives more weight to recent periods. Prophet handles gaps and sudden shifts well. In the model comparison (Pro plan) you can compare ARIMA, Exponential Smoothing and Prophet on the last periods of your own data and see which forecasts best.
Archive: your three newest analyses are in the sidebar. The Free plan has no archive, so older analyses are not shown. The Student plan keeps your last 5 archived analyses, the Starter plan your last 10, the Analyst plan your last 15 and the Pro plan your last 100. Older ones are hidden, not deleted, so they come back if you upgrade.
Reports with your own logo (Pro plan): in Export, choose Your logo on reports, add a PNG, JPG or SVG logo, or just your company name. Your Excel and PDF reports then show your logo (or name, never both) with "Powered by StatEVN" next to it. Without it they show the StatEVN logo.
- Actual vs predicted: each dot is a row. The closer to the dashed line, the better.
- Errors vs predicted (Analyst and Pro): errors should scatter evenly around zero. A curve or a funnel shape means the model is missing something.
- How big the errors are (Analyst and Pro): a symmetric hump centred on zero is what you want.
- Strength of evidence: a longer bar means a more reliable effect. Grey bars are not significant.
- ROC curve (yes/no models, Analyst and Pro): the further it bends towards the upper left, the better the model separates the two outcomes.
- Effect curve (boosted models): what the model predicts as one factor changes.
9. Reliability warnings
A high R² does not always mean a good model. StatEVN checks for situations where a flattering number hides a problem, and shows a warning when it finds one:
- a factor that is almost a copy of the outcome, or an outcome that is a formula of the factors (a total and its parts);
- the same item repeated on many rows, which makes the test score look better than it is;
- strongly uneven variance;
- factors that are almost copies of each other.
The warnings are signs to check, not proof of a mistake. The message tells you what looks wrong and what to try.
10. What-if simulation
On the Starter (20 simulations a month), Analyst and Pro plans you can ask "what happens if…" after a regression. Choose a factor and one of three ways to change it:
- Change by % moves the factor by a percentage of its typical value.
- Add or subtract units changes it by a fixed amount.
- Set exactly to sets it to a value you choose.
StatEVN shows the typical (average) value of the factor next to your input, runs thousands of simulations that account for the model's uncertainty, and reports the likely change in the outcome, including the chance it goes up.
11. Exporting
- CSV for each table (interpretation, coefficients, model comparison, correlations) and JSON with all the data are available on every plan.
- Excel report is available on the Starter, Analyst and Pro plans, the PDF report on the Analyst and Pro plans. The PDF is a formatted report you can send to a colleague, and the Excel report has one sheet per section.
12. Plans and limits
| Plan | Analyses per month | Rows per file | Includes |
|---|---|---|---|
| Free · Regular | 10 | 5,000 | Amateur mode, core models, CSV and JSON export |
| Free · Student | 30 | 10,000 | As Regular, for verified students and academics |
| Starter | 40 | 25,000 | Core models plus GLMs and Exponential Smoothing, result charts, Excel reports, pivot unpivoting, A/B test tool, 20 what-if simulations a month |
| Analyst | 100 | 100,000 | All models, Advanced mode, all charts, Excel and PDF reports, pivot unpivoting, A/B test tool, unlimited what-if simulation |
| Pro | 500 | 500,000 | Everything in Analyst, plus priority compute |
Only analyses that complete successfully count towards your monthly limit. Paid plans are not open for purchase yet and details may change before launch. See the pricing section for the current overview.
13. Your data
Your files are used only to produce your results and are not saved. StatEVN does not send your data to an AI service: the explanations are written by StatEVN's own rules from the statistical results. See the Privacy Policy for details.
14. Common questions
Why did StatEVN use a different model than I expected?
It chooses the model that fits your data and goal, and shows the others it tested on the same data.
Why does my result carry a warning?
The reliability check found something that can make the numbers look better than they really are. Open the warning for the reason and suggestion.
Why was a column or some rows removed?
Open the Data cleaning summary. It lists dropped columns (mostly empty), removed rows (too little information) and columns left out (too many different values).
Can I use StatEVN to forecast?
Yes. When your data is a series over time, StatEVN can fit ARIMA (or SARIMAX when you add external variables) and show a forecast for the next periods. If your goal uses the word "predict", StatEVN asks whether you want to understand the effect of other variables (regression) or forecast a variable over time, and opens the matching variable selection.
How do I check the result of an A/B test?
The A/B test tool is part of the Starter, Analyst and Pro plans. Each A/B test you run counts as one analysis in your monthly total, the same as a regular analysis; the sample-size planner does not count. Open A/B test in the top bar. You can enter totals (visitors and conversions per version, or number, average and standard deviation for a number such as revenue), upload a file with one row per visitor, or plan how many visitors a test needs before it starts. StatEVN shows each version's rate with a 95% confidence interval, the difference and the relative lift with their intervals, a p-value from the right test, and a plain-English verdict. With several variants each is compared with the control and the p-values are adjusted. It also warns when the groups are not the size an even split would give, and tells you the smallest change your test could reliably detect.
How do I test many segments at once, such as every country and product?
Pro plan. Open A/B test, then Segments table. Copy your table from Excel or Google Sheets (header row included) and paste it, or upload the file. You need one row per segment with the visitors and the purchases for the old and the new version. StatEVN guesses the columns and shows the first row in plain words so you can check them. Add the price columns to get the break-even (how many more purchases a price cut needs just to keep revenue level) and the revenue columns to compare revenue per visitor. You get combined results by product, by country or overall, and a table with a verdict for every segment. The p-values are adjusted for the number of segments, because testing many rows always produces some lucky ones. Segments with few purchases show "Can't tell yet", with the smallest change that segment could have detected.
Why does it say "No clear winner yet"?
The difference you see is no bigger than what luck alone can produce with this many visitors. It does not prove the versions are the same. Check the smallest detectable change in the result: if it is larger than the change you hoped for, you need more traffic. Decide the sample size before the test and read the result once, because stopping the moment the p-value looks good produces false winners.
My file is too big for my plan.
Use a smaller file, group the data first, or move to a plan with a higher row limit.
I have another question.
Use the feedback form. Leave your email there if you would like a reply.