Crafte Sessions analysis 2026 09 29 2026 10 29
sessions_analysis-2026-09-29-2026-10-29.csv · 2,246 rows · 15 columns · definition v3 saved 29 Oct 2026 14:08
Rows
2,246
all used
Entities
2,014
median 1 row each, max 9
From
09 Oct 2026
to 29 Oct 2026
Over time
20 days of dataRows per day across 20 days.
- Rows per day across 20 days.
Rows by day
Per person
8% of people come back; the median person has 1 row.
- 8% of people come back; the median person has 1 row.
- This view is per person (entity), not per row: one-off visitors and regulars count once each.
- Outcome on the first row: of people who ever had the outcome, the share who had it on their very first row. High means most decide in one visit; low means it builds up over visits.
- Chance on each row: of all first rows, second rows and so on, the share where the outcome happened. Rising means people who come back are more likely to act each time.
People
2,014
one row each: 92%
Came back
8%
median 1 row, 90% have 1 or fewer
Rows per person
| Rows | People | |
|---|---|---|
| 1 | 1,849 | |
| 2 | 125 | |
| 3 | 25 | |
| 4 | 11 | |
| 5 | 1 | |
| 6 | 1 | |
| 8 | 1 | |
| 9 | 1 |
Earlier vs later people
1,828 earlier rows, 418 laterlandingPage shifted a lot between earlier and later people. Later people came back 4% of the time against 9%: they've had less time to.
- landingPage shifted a lot between earlier and later people. Later people came back 4% of the time against 9%: they've had less time to.
- The people are split the way the models split them (the dataset's train/test split): earlier people train, later people test. If the two groups differ, a model can score well in training and badly on the test.
- PSI (population stability index) measures how much a column's mix changed: under 0.1 stable, 0.1 to 0.25 some shift, above 0.25 a big shift.
- For numbers, the median and average are compared too; for categories, the values whose share moved most.
- Came back: later people have had less time to return, so anything built on history (rows so far, tenure) will look different for them. That's not a fault in the data, but it lowers test scores.
People: earlier / later
1,611 / 403
later = first seen after 25 Oct 2026
Came back
9% → 4%
earlier → later people
| Column | Shift | PSI | Earlier → later |
|---|---|---|---|
| landingPage | big shift | 0.38 | / 86% → 68%; /product/79-chocolate-fruit-ca 3% → 9%; /product/75-christmas-cake-box 1% → 3% |
| newVisitor | some shift | 0.11 | 1 88% → 96%; 0 12% → 4% |
| source | stable | 0.06 | Facebook 95% → 99%; Direct 4% → 1%; Crafte 1% → 0% |
| device | stable | 0.04 | mobile 74% → 81%; desktop 13% → 7%; tablet 14% → 12% |
| pages | stable | 0.01 | median 1 → 1 (average 1.69 → 1.68) |
| minutes | stable | 0.00 | median 0 → 0 (average 9.71 → 14.4) |
Data quality
Every column, things to check firstNothing to do: every flag is already handled, expected, or on a column that isn't used.
- Nothing to do: every flag is already handled, expected, or on a column that isn't used.
- Explore only reports: nothing is changed. Fixes are made on the Columns tab, then the features are rebuilt.
- To do: worth fixing. Handled: your column settings already deal with it. Expected: true of this kind of column (an ID has many values, a rare outcome is mostly 0). Not used: the column isn't used, so it doesn't matter.
- 0 duplicate rows; 2,031 rows with an empty value in a column that has values.
- Top value: the share of rows holding the most common value. Near 100% leaves little to learn from.
- Outliers: values beyond three times the middle 50% range from the quartiles. They can be real (a big order) or errors; a log transform or capping on the Columns tab stops them dominating.
- Skewed: a long tail on one side. Models and segments see it better after a log transform.
- Case clashes: 'Facebook' and 'facebook ' counted as different. Tick 'Ignore case and spaces' on Columns.
| Column | Empty | Distinct | Top value | To check |
|---|---|---|---|---|
| minutes feature | 0% | 111 | 83% | Handled Skewed: Very skewed (skew 7.1). Handled: transform is log on Columns. Expected Mostly zeros: 83% zeros. Real: most rows have none. Keep it. |
| pages feature | 0% | 24 | 78% | Handled Skewed: Very skewed (skew 10.2). Handled: transform is log on Columns. |
| actions feature | 90% | 58 | 32% | Expected Many empty: 90% empty. Empty means none: counted as 0. |
| browserID entity id | 0% | 2,014 | 0% | Expected Many values: 2,014 different values, most of them rare. Expected for an ID or time column. |
| ordered outcome | 0% | 2 | 97% | Expected One value dominates: 97% of rows are 0. Expected: a rare outcome. |
| source feature | 0% | 7 | 96% | Expected One value dominates: 96% of rows are Facebook. Real, little to learn from most rows; the rarer values can still matter (see Each column vs outcome). |
| customerID ignore | 100% | - | - | Not used Empty: Every value is empty. Not used (role Ignore). |
| addedToCart ignore | 0% | 2 | 89% | fine |
| device feature | 0% | 3 | 75% | fine |
| firstSeen timestamp | 0% | 2,210 | 0% | fine |
| landingPage feature | 0% | 53 | 83% | fine |
| lastSeen feature | 0% | 2,215 | 0% | fine |
| newVisitor feature | 0% | 2 | 89% | fine |
| reachedCheckout ignore | 0% | 2 | 92% | fine |
| sessionDate feature | 0% | 21 | 8% | fine |
Distributions
Each number column's shape, with the standard distribution that fits it best.
- Each number column's shape, with the standard distribution that fits it best.
- Each candidate is fitted by maximum likelihood: the parameters that make the observed values most probable.
- AIC compares the fits: how well each explains the values, with a penalty per parameter. Lower is better; within 2 of the best is effectively as good, more than 10 behind is clearly worse.
- Whole numbers from 0 (counts) are compared as Poisson and negative binomial; other numbers as normal, log-normal, gamma, exponential and Weibull (the last four only when every value is above 0).
- Why it matters: a right-skewed shape (log-normal, gamma) says use a log transform; an over-dispersed count says a Poisson model will be over-confident.
minutes
2,246 valuesminutes is best described as Negative binomial (mean 10.6, dispersion r 0.0307: variance above the mean). Next: Poisson (+141,560 AIC).| Shape | AIC vs best | Fitted |
|---|---|---|
| Negative binomial | best | mean 10.6, dispersion r 0.0307: variance above the mean |
| Poisson | +141,560 | mean 10.6; variance should equal the mean |
Variance is 304.1× the mean, more spread than a Poisson count: use negative binomial (or Poisson with care) when modelling it. 83% zeros against 0% a Poisson would give: extra zeros, often a separate 'did anything at all' step.
pages
2,246 valuespages is best described as Negative binomial (mean 1.69, dispersion r 3.36: variance above the mean). Next: Poisson (+671 AIC).| Shape | AIC vs best | Fitted |
|---|---|---|
| Negative binomial | best | mean 1.69, dispersion r 3.36: variance above the mean |
| Poisson | +671 | mean 1.69; variance should equal the mean |
Variance is 3.7× the mean, more spread than a Poisson count: use negative binomial (or Poisson with care) when modelling it.