JEE Main · Method and backtest
How far to trust it
Tested on 2022–2026, the 30 chapters Lumi ranks highest held 53% of a shift's questions, against 37% for 30 picked at random. A plain all-time count of past papers did about as well (52%). Past papers tell you where marks have been; recency weighting adds little on top, and nothing here names the next shift's questions.
- Top 30 chapters, hit rate
- 93.1%
- random 74.4%
- Top 20 topics, hit rate
- 68%
- random 32%
- Top 100 ideas, hit rate
- 9.3%
- random 2.4%
- Chapter rank correlation
- 0.75
- all-time count 0.71
Sources
What went in
Each question comes from an official JEE Main Paper 1 question paper published by NTA (CBSE for 2017), and each answer from the final answer key. Lumi analyses a shift only when both exist. Of the 174 shifts held from 2017 to 2026, 108 are in: 9,285 questions, of which NTA dropped 44. No 2018 or 2019 shift is in this analysis. The coverage table lists every shift held and why it is or is not here.
Every question was tagged with its chapter, topic and the exact idea it tests, plus format, skill and difficulty. Of the 9,285 questions, 4,927 fit an idea in Lumi's JEE taxonomy closely and 2,823 fit it reasonably. 1,535 had no close idea; they count for their chapter but carry no topic or idea. No answer key needed an erratum. Difficulty is Lumi's rating, not NTA's.
Chapters follow NTA's 2026 syllabus. Chapters it no longer lists are kept, marked as out of the syllabus, because their older questions are still useful practice. Syllabus checked against Syllabus for JEE (Main) - 2026, Paper 1 (B.E./B.Tech.), NTA, linked from jeemain.nta.nic.in on 2026-10-04.
Backtest
Tested on years it had not seen
For each year from 2022 to 2026, Lumi ranked chapters, topics and ideas using only earlier years, then scored the ranking on every analysed shift of that year, next to three simpler ways of picking. The score is the share of a shift's 75 questions the top picks held.
History against guessing
23% vs 10%
The 20 topics asked most in earlier years, counted with no weighting at all, held 23% of the next shift. Twenty topics picked at random held 10%. Most of what past papers can tell you comes from simply counting them.
Weighting against counting
24% vs 23%
Letting recent years count more added 0.7 points for the top 20 topics, about 0.5 of a shift's 75 questions, under the one-question tie line. For the top 30 chapters it was 53% against 52%: a tie, within one question a shift. On exact ideas it was ahead (14% against 10% for the top 100), but 100 ideas still held only about 10.5 of 75 questions.
Four ways to pick, side by side
Share of a shift's questions held by each method's top picks, averaged over the analysed shifts of 2022–2026. A random pick is the expected share for that many chosen blind. A gap under one question of a 75-question shift counts as a tie. Against the all-time count, the weighting was ahead for 0 of 3 chapter lists, 2 of 3 topic lists and 3 of 3 idea lists.
- Top 15 chaptersTie · +0.8 Q
- Weighted estimate30%
- All-time count29%
- Last year only30%
- Random pick19%
- Top 30 chaptersTie · +0.7 Q
- Weighted estimate53%
- All-time count52%
- Last year only52%
- Random pick37%
- Top 45 chaptersTie · +0.4 Q
- Weighted estimate73%
- All-time count72%
- Last year only71%
- Random pick56%
- Top 20 topicsTie · +0.5 Q
- Weighted estimate24%
- All-time count23%
- Last year only24%
- Random pick10%
- Top 40 topicsAhead · +1.4 Q
- Weighted estimate41%
- All-time count39%
- Last year only41%
- Random pick21%
- Top 60 topicsAhead · +1.9 Q
- Weighted estimate54%
- All-time count52%
- Last year only54%
- Random pick31%
- Top 50 ideasAhead · +2.9 Q
- Weighted estimate9%
- All-time count5%
- Last year only9%
- Random pick2%
- Top 100 ideasAhead · +3.0 Q
- Weighted estimate14%
- All-time count10%
- Last year only14%
- Random pick3%
- Top 200 ideasAhead · +1.2 Q
- Weighted estimate23%
- All-time count21%
- Last year only21%
- Random pick7%
| List | Weighted estimate | All-time count | Last year only | Random pick | Weighting vs count |
|---|---|---|---|---|---|
| Top 15 chapters | 30% | 29% | 30% | 19% | Tie · +0.8 Q |
| Top 30 chapters | 53% | 52% | 52% | 37% | Tie · +0.7 Q |
| Top 45 chapters | 73% | 72% | 71% | 56% | Tie · +0.4 Q |
| Top 20 topics | 24% | 23% | 24% | 10% | Tie · +0.5 Q |
| Top 40 topics | 41% | 39% | 41% | 21% | Ahead · +1.4 Q |
| Top 60 topics | 54% | 52% | 54% | 31% | Ahead · +1.9 Q |
| Top 50 ideas | 9% | 5% | 9% | 2% | Ahead · +2.9 Q |
| Top 100 ideas | 14% | 10% | 14% | 3% | Ahead · +3.0 Q |
| Top 200 ideas | 23% | 21% | 21% | 7% | Ahead · +1.2 Q |
Year by year
Each line is one graded year, averaged over its analysed shifts (24 in 2022, 12 in 2023, 18 in 2024, 8 in 2025, 9 in 2026). Across these 15 comparisons the weighting was ahead of the all-time count 6 times, tied 9 and behind 0.
Top 30 chapters
Against the all-time count: ahead in 1, tied in 4, behind in 0 of 5 years.
- 2022Tie · +0.8 Q
- 2023Tie · +0.4 Q
- 2024Ahead · +1.6 Q
- 2025Tie · +0.7 Q
- 2026Tie · −0.2 Q
Top 20 topics
Against the all-time count: ahead in 2, tied in 3, behind in 0 of 5 years.
- 2022Ahead · +1.1 Q
- 2023Ahead · +2.2 Q
- 2024Tie · −1.0 Q
- 2025Tie · +0.2 Q
- 2026Tie · +0.3 Q
Top 100 ideas
Against the all-time count: ahead in 3, tied in 2, behind in 0 of 5 years.
- 2022Tie · 0.0 Q
- 2023Tie · +1.0 Q
- 2024Ahead · +3.9 Q
- 2025Ahead · +6.2 Q
- 2026Ahead · +3.9 Q
- Weighted estimate
- All-time count
- Last year only
- Random pick
| Rank correlation | Average miss | |
|---|---|---|
| Weighted estimate | 0.75 | 0.26 |
| All-time count | 0.71 | 0.28 |
| Last year only | 0.69 | 0.27 |
| Same weight for every chapter | – | 0.39 |
Rank correlation with the next year's chapter mix, and the average miss in questions per 25 a subject.
| 2022 | 2023 | 2024 | 2025 | 2026 | |
|---|---|---|---|---|---|
| Weighted estimate | 95% | 93% | 96% | 91% | 91% |
| All-time count | 94% | 93% | 95% | 90% | 92% |
| Last year only | 94% | 91% | 91% | 87% | 93% |
| Random pick | 83% | 83% | 73% | 66% | 67% |
Share of the 30 chapters that came up in a shift. Almost every big chapter comes up, so even a random 30 scores high.
| Hit rate | Covered | |
|---|---|---|
| Weighted estimate | 93.1% | 53.0% |
| All-time count | 92.7% | 52.1% |
| Last year only | 91.2% | 52.0% |
| Random pick | 74.4% | 37.0% |
| Hit rate | Covered | |
|---|---|---|
| Weighted estimate | 67.6% | 23.5% |
| All-time count | 65.4% | 22.8% |
| Last year only | 66.7% | 23.6% |
| Random pick | 31.8% | 10.2% |
| Hit rate | Covered | |
|---|---|---|
| Weighted estimate | 9.3% | 14.0% |
| All-time count | 6.7% | 10.0% |
| Last year only | 9.3% | 13.6% |
| Random pick | 2.4% | 3.4% |
Sensitivity
Does the answer depend on the settings?
The chapter estimate, rebuilt four other ways and compared with the one Lumi uses.
| Change | Rank correlation | Same top 30 | Largest move | Moved most |
|---|---|---|---|---|
| Weight years by how many shifts they have | 0.991 | 30 of 30 | 0.13 Q | Ray Optics and Optical Instruments |
| Leave out 2017 (one offline paper) | 0.998 | 30 of 30 | 0.07 Q | Haloalkanes and Haloarenes |
| Let old years fade half as fast (half-life 6 years) | 0.997 | 29 of 30 | 0.12 Q | Application of Derivatives |
| Ignore the 2026 syllabus cuts | 0.993 | 27 of 30 | 0.36 Q | The s-Block Elements |
Every variant ranks the chapters almost the same way (rank correlation 0.991 or higher). The chapter order on these pages does not hang on one choice of settings. Ignoring the syllabus cuts moves the most (The s-Block Elements), because it puts dropped chapters back in. "Largest move" is in questions per 25 a subject.
Within a year
Do shifts in one session look alike?
Each 2024, 2025 and 2026 shift was held out in turn and ranked from the rest. Averages over the three years.
| Built from | Hit rate | Covered |
|---|---|---|
| Earlier years only | 58% | 42% |
| Earlier years + the same year's other shifts | 62% | 45% |
| The same year's other shifts only | 62% | 45% |
Random pick: 30% hit rate.
| Built from | Chapter rank corr. | Top 100 ideas hit |
|---|---|---|
| Earlier years only | 0.60 | 10.8% |
| Earlier years + the same year's other shifts | 0.66 | 11.3% |
| The same year's other shifts only | 0.66 | 9.0% |
Random pick: 2.0% for ideas.
Shifts in the same session share their topics far more than their exact ideas. Adding the year's other shifts to the history lifts the topic hit rate from 58% to 62%, and barely moves ideas (10.8% to 11.3%). It does not reveal questions. Over every pair of analysed shifts, the chapter mix of two shifts in the same year correlates at 0.38 (901 pairs), against 0.30 for shifts in different years (4877 pairs).
Calibration
When it says 60%, does it happen 60% of the time?
Every stated chance, grouped, against how often the unit actually came up in the shift.
| Stated chance | Came up | Cases |
|---|---|---|
| 0–5% | 11% | 36 |
| 5–10% | 9% | 35 |
| 20–30% | 35% | 162 |
| 30–50% | 55% | 1,330 |
| 50–70% | 83% | 3,042 |
| 70–90% | 94% | 1,125 |
| 90–100% | 100% | 21 |
| Stated chance | Came up | Cases |
|---|---|---|
| 0–5% | 9% | 845 |
| 5–10% | 13% | 1,535 |
| 10–20% | 19% | 2,990 |
| 20–30% | 30% | 3,309 |
| 30–50% | 47% | 4,271 |
| 50–70% | 71% | 888 |
| 70–90% | 89% | 18 |
| Stated chance | Came up | Cases |
|---|---|---|
| 0–5% | 2% | 1,28,087 |
| 5–10% | 6% | 15,598 |
| 10–20% | 9% | 5,690 |
| 20–30% | 16% | 309 |
| 30–50% | 27% | 33 |
The stated chances run low in every band for chapters, low in every band for topics and high in every band for exact ideas (low: the unit came up more often than stated). Read the chances as a ranking, not as exact odds.
See the shifts by year, starting with JEE Main 2026.
Priority bands
How a chapter gets its band
Within each subject, chapters in NTA's 2026 syllabus are ranked by their average questions a shift over all 8 years analysed (history, not the model). Going down that list, the chapters that together held the first half of the subject's questions are Must do; the next ones, up to 80%, are Should do; the rest are Lower priority.
Bands use history rather than the weighted estimate because the backtest above shows weighting adds little over plain counting, and history is the part Lumi can state as fact. Bands are set within each subject, because every shift asks the same number of questions in each. Study hours on chapter pages are what the study plan gives each chapter; they are a planning model, not timings of real students.
- Must do28chapters · First 50% of a subject’s questionsPhysics 11 · Chemistry 9 · Maths 8
- Should do22chapters · 50% to 80%Physics 10 · Chemistry 6 · Maths 6
- Lower priority20chapters · Last 20%Physics 7 · Chemistry 6 · Maths 7
11 chapters are out of NTA’s 2026 syllabus and get no band.
Limits
What this cannot tell you
- 108 of 174 shifts is a sample. Lumi analyses a shift only when the full paper and final key are both available, so some years carry more shifts than others (2017: 1 of 3). Year-by-year figures are shares or per-shift averages for that reason, but a year with few shifts still rests on few papers.
- Tagging was done and checked by AI. No person re-tagged a sample independently, so there is no measured agreement rate. Of 9,285 questions, 1,535 had no close idea in Lumi’s map; they count for their chapter and topic but not as an idea.
- Shifts in one year are not independent. The chapter mix of two shifts in the same year correlates at 0.38, against 0.30 for shifts in different years. Repeat figures across years compare against earlier years only.
- Early years have little history. “New idea” in 2020 means new since 2017 only, and 2017 was a single offline paper. The share of new ideas looks higher in early years partly for that reason.
- The syllabus changed. 11 chapters are out of the 2026 syllabus and 6 only partly in. Patterns from older shifts count for less in chapters that changed.
- 8 years is a short history. One surprise shift moves a chapter’s average. Treat weightage as a guide to effort, never as a forecast of the paper.
Questions