I Should Be Looking for a Job

The polite thing to do during a job search is network, polish the resume, and answer recruiter email promptly. I have done some of that, I promise. But I’ve also spent a healthy chunk of this fall learning American mah jongg instead — and because I have never met a file format I didn’t want to parse, I ended up turning fourteen years of the National Mah Jongg League’s annual cards into statistics.

If you don’t play, here’s the short version: American mah jongg is its own thing, distinct from the riichi and Hong Kong styles most of the world plays. The National Mah Jongg League (NMJL) has been publishing an annual card of winning hands since the 1930s, and the card is the game — when the new one arrives each spring, the whole landscape of valid hands changes and every player re-learns it. The card is copyrighted, and selling it is the league’s business, so this post reproduces no complete hands — aggregate statistics only, apart from a few tiny notation fragments down in the transcription-quirks section where the digitization itself has typos. If you play, buy the real card from nationalmahjonggleague.org. Consider that my good deed for the day.

The full analysis — both CSVs, the method, the caveats — lives in a gist, which I’m keeping around as the permanent record for this little project. This post is the guided tour. If you want to load the data into a spreadsheet and make your own charts, the gist is your playground.

The Headline Numbers

Fourteen consecutive years of cards, 2013 through 2026, 749 named hands in total. Percentages are of each year’s named hands as digitized (51 to 55 per year — there’s a unit-of-analysis caveat later, and it matters):

Year Hands Dragons Winds Flowers Concealed Req. joker No joker
2013 52 28.8% 15.4% 42.3% 26.9% 7.7% 13.5%
2014 52 32.7% 15.4% 42.3% 26.9% 7.7% 13.5%
2015 51 31.4% 15.7% 47.1% 25.5% 7.8% 13.7%
2016 52 30.8% 19.2% 46.2% 28.8% 7.7% 13.5%
2017 52 28.8% 13.5% 40.4% 26.9% 7.7% 13.5%
2018 54 37.0% 18.5% 48.1% 25.9% 7.4% 13.0%
2019 53 32.1% 17.0% 45.3% 26.4% 7.5% 13.2%
2020 54 33.3% 18.5% 42.6% 25.9% 7.4% 13.0%
2021 54 33.3% 20.4% 44.4% 22.2% 7.4% 11.1%
2022 55 30.9% 16.4% 41.8% 23.6% 7.3% 10.9%
2023 55 32.7% 16.4% 45.5% 23.6% 5.5% 10.9%
2024 55 30.9% 18.2% 45.5% 21.8% 7.3% 10.9%
2025 55 38.2% 18.2% 49.1% 23.6% 5.5% 10.9%
2026 55 34.5% 21.8% 40.0% 21.8% 5.5% 10.9%

Pooled across all 749 hands: 32.4% use dragons, 17.5% use winds, and 44.4% use flowers; 25.0% are concealed; 6.9% require a joker; and 12.3% cannot contain one. Expand the multi-option hands and you get 1,110 enumerated patterns — pattern-level percentages are in the CSV.

What the Numbers Say

Dragons, Winds, and Flowers

  • Dragons (29-38%) never dip low because the year-hand section uses soap — the white dragon — as its zeros, which makes that section 100% dragons every single year. Add the Winds & Dragons section and the scattered dragon kongs elsewhere, and dragons are guaranteed a healthy share of any year’s card. The 2025 peak (38.2%) checks out against independent analysis of the printed card: iLoveMahj noted dragons were better represented in 2025 than 2024 (38.2% vs 30.9% here).
  • Winds (13-22%) are the rarest of the three and live mostly in the Winds & Dragons section. 2026 is the windiest card of the series at 21.8%, with winds in six of nine sections.
  • Flowers (40-49%) are the workhorse. Pairs, pungs, and kongs of flowers appear across every section, which is why every mahj teacher tells you never to pass flowers in the Charleston. The numbers back up the advice.

Hands That Require a Joker (5.5-7.8% per year)

Across all 14 years and 749 hands, not one hand prints a joker in its pattern. The digitization’s notation supports printed jokers — the token J exists — and no NMJL card file from 2013 to 2026 uses it, not once. Joker-requirement is always structural: any quint or sextet of a tile with only four copies in the 152-tile set. That turns out to be exactly the Quints section, every year — four hands (7.7%) through 2022, three hands (5.5%) since 2023 when the section slimmed down. One 2026 nuance: the new flower sextet hand (six of eight flowers) is the one super-group hand that does not need a joker.

Hands That Cannot Contain a Joker (10.9-13.7% per year)

Jokers may only substitute in groups of three or more identical tiles — pungs, kongs, quints, sextets — so a hand with no such group can never hold one. That is exactly the Singles & Pairs section, every year, no exceptions: seven hands (~13.5%) through 2020, six (~11%) since 2021. Note that concealed does not mean no-joker: concealed hands with pungs happily use jokers in those groups.

Which leaves the economy this post is named after: roughly 7% of hands force you to have jokers, ~12% forbid them, and the ~81% in between are where joker strategy actually lives. Fourteen years of cards, and that split barely wobbles.

How I Counted

  • Source: the card files (am_card_nmjl_2013.jq through am_card_nmjl_2025.jq, plus am_card_nmjl_2026.majs) in the Riichi Advanced repository (AGPL-3.0), an open-source mahjong client that digitizes each year’s named hands, point values, concealed flags, and tile patterns as playable rulesets. Fourteen years of card files sitting in a public repo — you can see how a person who parses things for a living might lose an afternoon to that. (And no, .jq is not the jq tool, though I enjoyed the coincidence.)
  • Notation semantics verified against the engine’s own parser (lib/riichi_advanced/game/game_state/american.ex): digits are literal tiles (0 is soap, the white dragon); X-prefixed tokens are relative offsets (flexible numbers); D plus a suit letter is a suited dragon; F is flowers; N/E/W/S are winds; and Z is a wind wildcard (e.g. ZZZZ is a kong of any wind).
  • Every hand was parsed, its | alternatives expanded, and each expanded variant validated to sum to exactly 14 tiles. Fifteen hands got flagged — see the quirks below.
  • Joker-eligibility follows the engine’s own logic: groups of three or more identical tiles are joker-usable; singles, pairs, and mixed groups (like NEWS) are tagged no-joker.

Definitions

  • Hand — one named hand in the digitization (a “hand family”); multi-option hands count once (“NNNN EEE WWW SSSS or NNN EEEE WWWW SSS” is one hand with two patterns).
  • Uses dragons — the pattern contains a dragon group or a soap-as-zero tile in at least one variant.
  • Requires joker — every variant contains a group of five or more identical non-flower tiles, impossible to build from the four-copy tile supply without jokers (flowers excepted; there are eight).
  • No joker allowed — no variant contains any group of three or more identical tiles.
  • Concealed — flagged concealed in the digitization (has_no_call_named).

Caveats

Every dataset has fine print. Mine has more than most, which is what happens when the source is a fan transcription of a copyrighted card:

  • Unit of analysis. The 51-55 “hands” per year are hand families as digitized. The printed 2025 card lists 71 lines, and fully expanded playable hands number 1,002 by iLoveMahj’s count — percentages shift a few points depending on the unit. The main CSV includes pattern-level percentages (over the 73-104 enumerated patterns per year) so the unit effect is visible.
  • Unofficial digitization. This is a fan transcription of a copyrighted card, itself unlicensed. Treat the numbers as derived data and cross-check anything that matters against the printed card.
  • Z means winds, not jokers. In this notation Z is a wildcard wind. An early draft of this analysis misread those tokens as printed jokers; the corrected reading is what is documented here, and the zero-printed-joker finding reflects it. Read the parser, not just the data.
  • Cross-validation. Against iLoveMahj’s 2025 card analysis: dragons up in 2025 vs 2024 (38.2% vs 30.9% here, directionally matching their note); no flower/dragon/wind quints on the 2025 card (matching); flowers in 52.2% of playable hands there vs 49.1% of hand families here (different units, same ballpark).

Transcription Quirks

The 14-tile validation flagged fifteen hands across fourteen years. None of them change the headline statistics, but as someone who has typoed a config file or two in his life, I present them with sympathy:

  • 2013 — a 33ea typo in 13579 #4 (likely 333a); Winds & Dragons #3/#4 ask for six of one suit (111a 111a, presumably two suits; the same hand family recurs in 2015 — those two extra flags are part of the fifteen).
  • 2014 — year hand #1 wants five Souths (wind-letter typo); 369 #3/#5 put two kongs or pungs of the same tile in one suit (eight or six copies).
  • 2016 — Quints #1 is missing spaces (XX0aX1aX2a); 369 #5 repeats the same-suit duplication quirk.
  • 2021 — year hand #3 wants five 2s of one suit (suspected suit typo).
  • 2022 — Consecutive Run #2 is 12 tiles (two short).
  • 2026 — Consecutive Run #5’s second option is 13 tiles (likely XX1b should read XXX1b).
  • 2020 #2 and 2022 #1 — kongs of reds and greens together: legal hands, flagged only by the validator’s conservative tile-pool check.

Go Play With the Data

The gist holds both CSVs:

  • nmjl_card_stats_2013_2026.csv — per-year statistics, 21 columns. The hand_* columns are shares and counts of named hands; the pattern_* columns are shares of enumerated patterns; the *_n columns are counts.
  • nmjl_sections_2013_2026.csv — per-section, per-year breakdown (135 rows). Counts are hands in that section having each property. Section names are as they appear in the digitized files, including the rotating “gimmick” section (Seven Hands, Eleven Hands, Lucky 13, Addition Hands, Lucky Sevens — absent after 2024).

Credit where it is due: the data source is Riichi Advanced (AGPL-3.0), an open-source mahjong client with digitized NMJL cards for 2013-2026. The cross-checks come from iLoveMahj’s 2025 card analysis. And the cards themselves belong to the National Mah Jongg League — buy yours, and whatever you do, don’t pass your flowers in the Charleston.

If a recruiter asks, this was professional development in data analysis.