WWE Champions·Growth diagnosis & proposal·August 2026
50 million people installed it.
The one-star reviewers
still call it fun.
Across 2,945 player reviews, the word “fun” appears at the same rate in one-star reviews as in five-star reviews — while “money” is the most-used word in the whole corpus. The loop is loved and the transaction is hated. That is a first-purchase conversion problem, and it is the one factor of ARPDAU nobody on this title has ever worked on. Here's what the data says, how I know it, what it's worth, and what I'd build.
Start with the case → Go to the proposal → Read the specs → Play the prototype → Sources & method →
A correction · 8 August
An earlier version of this analysis said the League 2 featured shelf held exactly one offer,
at $99.99, and that the entry price band was empty. Re-reading my own source screenshot
(IMG_0821) to rebuild the prototype, that is wrong. Beneath the $99.99 League Essentials
pack, the same screen carries a 2-Star Gold Andre Starter Pack at $14.99, marked
“One Time Offer”, on a 05h 42m countdown.
What changed: move 02 now adds two rungs beneath an existing pack rather than building a ladder from nothing; its modelled lift drops from +0.55pp to +0.35pp, and the programme's base case falls from $418k to $367k a year and +2.2% ARPDAU. What didn't: nothing sits between free and $14.99 — three to five times the conventional entry price for the genre — and the cheapest way into the game is a countdown timer aimed at somebody who installed this afternoon. Two prototype captions were wrong too, and are fixed; both are listed in Act 09.
The argument, in six answers
Grouped into three topics · the nine acts that follow are the evidence
The diagnosis
Evidence in Acts 01–04
What is the goal?
Raise ARPDAU — which means raising the share of players who pay at all.
ARPDAU has two factors: (payers ÷ DAU) × ARPPU. On a 9.5-year-old title with 50M+ installs, ARPPU is already close to its ceiling and the review corpus prices the cost of pushing it further. Payer rate is the factor nobody has worked on — and it is set in a player's first week, which is the week this game spends showing them a $99.99 pack.
Who is the customer?
The early-game spender, across their first three purchases.
Not “the new installer” — that's the group, not the cohort. P0→P1 is two to five sessions in with a roster they care about and a wall in front of them. P1→P2 has bought once and is the highest-propensity group in the game. P2→P3 is where the habit either forms or lapses. All three are blocked by the same thing: the cheapest item on the League 2 shelf is a $14.99 starter pack on a six-hour countdown, under a $99.99 headline pack — and the first losable match ends in a “rescue your Superstar for 100 cash” prompt.
Why optimise this cohort?
Because the loop is loved and the transaction is hated.
Across 2,945 reviews, the word “fun” appears at the same rate in 1–2★ reviews as in 4–5★ — a skew of −1.1pp. “Money” skews +26.7pp and is the most-used word in the entire corpus; trust in the odds is the lowest-rated theme at 2.06★. People who don't care about spending don't talk about money. That is high intent meeting a hostile transaction — a conversion problem, and conversion problems are cheaper to fix than desire problems.
How these three fit together
These are one chain, and the order matters. Decompose the metric and one factor is defended and one is untouched — so the brief is more payers, not bigger ones. That points at exactly one cohort: the players standing at the first purchase with nothing in front of them they can afford. The corpus then explains why they don't convert. Affection survives the one-star rating; what replaces it is a vocabulary of money.
An earlier version of this analysis chose “the brand-new installer” and ranked its bets on retention. That answer wasn't wrong so much as underspecified — retaining installs grows DAU, which is the denominator. Same ten minutes, same screens, different success metric. Act 06 shows exactly what that difference is worth.
Act 01 · Where the game is in its life →
Act 02 · How I looked, and what I couldn't see →
Act 03 · What 2,945 players actually said →
Act 04 · Five pain points, in players' own words →
The programme
Evidence in Act 05
What can we actually do?
One programme, three moves. Not a menu.
01 · Earn the ask — no monetisation surface before a win, so the first thing the game asks you to buy stops being a rescue from a defeat. 02 · Build the rung — two deterministic rungs beneath the existing $14.99 starter pack, which becomes the top of the entry ladder rather than its floor. 03 · Make it safe to buy — published odds and dupe-to-shard, scoped to the entry band so no existing drop rate moves. Each one unblocks the next, and none works alone.
Why in this order — and what we're parking
- 02 without 01. A great offer in front of someone who hasn't won yet.
- 02 without 03. They convert once and never twice.
- 01 + 03 without 02. A clean shop whose cheapest item is still $14.99.
Parked, with the condition each is waiting on. Rookie Crew — a good D30 play, but it retains rather than converts, so it builds the denominator; ships once a payer base exists, still gated on the faction correlate. Mid-game catch-up track — blocked on faucet-vs-sink data, and also a denominator play. Tentpole win-back — sequenced after move 02 so returners land on a real ladder. Deeper whale extraction — rejected; no headroom, and the corpus prices the goodwill cost.
What it's worth
Evidence in Acts 06–07
Why does it matter?
Converting a non-payer buys you retention. Retaining a non-payer does not buy you a payer.
That asymmetry orders the whole portfolio, and it's where the downstream D7 and D30 effect comes from. Base case for the programme: — a year, moving ever-pay conversion from 2.00% to — and ARPDAU by —. One caveat stated plainly: payers retaining better is correlational until it's matched on session count, which is why the downstream retention line is the smallest of the three revenue components in the model.
Which improvement ranks best?
All three ship. The real ranking is programme vs. the retention portfolio.
Scored against ARPDAU, the previously-recommended retention leads are worth — a year and move the metric by — — because they grow the denominator. The programme is worth less money and moves the metric up. That trade is the recommendation, and it turns on one query: does a new player currently see anything under $10? The cheapest thing I could find is $14.99 on a six-hour countdown. If pull 2 finds something below that, move 02 is wrong and this collapses back to a retention case.
Why the bigger revenue number is the wrong one
Scored the previous model's own way — every added user monetising at the full average ARPDAU — the retention leads are worth — a year and move ARPDAU by —. Neutral. And neutral by construction: that model multiplies added user-days by average ARPDAU, so the ratio cannot move. The answer was fixed before the model ran.
The correction is one assumption: the users a retention bet adds are marginal survivors of the activation window, and they convert to payer below the average rate. Score them at half and the same bets read negative. There is no assumption under which a retention bet raises ARPDAU — that isn't a modelling choice, it's what arithmetic does when you add to a denominator. The programme is robust to the same correction because its gain comes from the numerator.
Act 06 · Open the model and change any input →
Act 07 · The four data pulls and the build order →
Where this game is in its life
WWE Champions is a match-3 RPG built on one of the strongest licences in entertainment. It is also nine and a half years old and well past its growth phase. That combination decides which growth levers are worth pulling.
What that means for strategy
A mature live game has two honest options: extract harder from the players who remain, or widen the base again. The review corpus below shows what the first costs in goodwill. The second is the brief: engage more people.
That reframes the growth job. On a title this age, whales are already monetised near their ceiling. The upside is in the players who install, look around, and leave before anything has a chance to hook them — and in the social structure that turns a solo player into a committed one.
It also settles how the base gets widened. A nine-year-old match-3 title buys its installs in the same auction as the biggest casual puzzle games on the platform, against budgets it cannot match. You don't out-bid that. You out-convert it. Which puts every point of first-session retention to work twice — once as retention, and again as a cut to what an install effectively costs.
The competitive set, and one number I withdrew
Those four numbers are the lifecycle picture, and each is a direct read from a store API rather than an estimate. A game nine and a half years old, with a 50-million install base, rated 4.36, and still under active development. That is a mature title carrying a very large lapsed population and a product people demonstrably liked.
The field it's ageing in
The instinct on a licensed title is to benchmark it against the other licensed titles. That's the wrong frame, and the store data says so: nothing in wrestling is taking this game's players. The only other licensed WWE live mobile game is a card battler on a different loop; WWE 2K26 is console; AEW's mobile entry is an idle promotion-sim. The competitive set that matters is the mechanic — the match-3 collection RPGs a churned player actually opens next.
Same API, same method as the four numbers above. Every figure in this table can be re-fetched in a single call.1
| Title | Publisher | Released | iOS rating | iOS ratings | Last update |
|---|---|---|---|---|---|
| WWE Champions match-3 RPG, WWE licence | Scopely | 25 Jan 2017 | 4.31 | 30,026 | 27 Jul 2026 |
| Empires & Puzzles match-3 RPG, no licence | Small Giant / Zynga | 1 Mar 2017 | 4.69 | 101,353 | 22 Jul 2026 |
| WWE SuperCard card battler, WWE licence | 2K Sports | 14 Aug 2014 | 4.77 | 78,326 | 3 Aug 2026 |
| MARVEL Puzzle Quest match-3 RPG, Marvel licence | 505 Go | 3 Oct 2013 | 4.58 | 27,508 | 29 Jul 2026 |
iTunes Lookup API, US storefront, fetched 6 Aug 2026. Rating counts are iOS-only and are not installs — they are a proxy for engaged population on one platform, nothing more.
Two things fall out of that table, and neither is the one you'd expect.
- The licence is not the problem. The other WWE game is rated 4.77 by 2.6× as many iOS raters. Wrestling fans will happily rate a WWE mobile game five stars — they're doing it next door. Whatever is costing this title goodwill, it isn't the audience or the IP.
- The closest peer is the same age and three times the size. Empires & Puzzles launched five weeks after WWE Champions, in the same genre, with no licence at all — and carries 3.4× the iOS ratings at 4.69. Same nine years, same mechanic, no IP advantage. That is the benchmark worth studying, not SuperCard.
And the long-run comparable is MARVEL Puzzle Quest. It is the same machine — licensed characters, match-3 core, a collection ladder with an escalating top tier — launched three years earlier, and still shipping updates in 2026. It is what this structure looks like at thirteen years old.
Which makes it the most useful question in this section: where is MPQ's revenue and install curve now, relative to its peak, and does WWE Champions sit on the same line? That comparison is one paid-tool query for you and unavailable to me — and it would either validate this whole document or reframe it.
What I could not verify — and did not build on
One paywall accounts for every gap in this act. Third-party trackers are widely reported to put WWE Champions well below its 2021 peak on revenue and installs, and they publish revenue estimates for every title in the table above — putting Empires & Puzzles in the hundreds of millions a year and MARVEL Puzzle Quest far below its own peak. All of it traces back to Sensor Tower pages that need a login, and I could not open a primary source. Blog aggregations of them aren't one either.
An earlier draft of this document led with a “60% revenue decline”. I've removed it — nothing here depends on a number I can't show you. So none of the argument in this act rests on revenue at all. It rests on release dates, ratings, install counts and update cadence, which are public and re-fetchable: the comparison is structural, not financial, and the financial half is a question for someone with the licence to a data tool.
I'm flagging that rather than dropping it silently, because if the decline is real it sharpens the case considerably — and you can see the true curve in about thirty seconds. That's a question for you, not a claim from me.
How I looked, and what I couldn't see
No internal telemetry. So the research had to be built from what is observable — and it had to be explicit about the difference between what players say and what the data proves.
The discipline. Every claim in this document is one of three things: a verified fact, a player-reported signal, or a named gap. Player complaints are treated as hypotheses to validate against funnel and cohort data — never as measurement.
Two limits worth stating before anyone finds them. Written reviews skew negative by construction: only 8.1% of raters write text, and angry players write. And onboarding complaints are just 3% of the corpus — expected, because players who churn on day one never come back to write a review. That absence is itself the argument for pulling the funnel data.
The three methods, and the six things I'd need from you
Three methods
- Played it. A new account to League 2, 42 screenshots of the activation window. Every claim below cites a specific screen, re-verified against the image.
- Read it at scale. 2,945 unique reviews — 474 iOS, 2,471 Android — sampled across newest, rating and helpfulness, then de-duplicated and analysed for word frequency, theme reach, and per-rating divergence.
- Anchored it. Google Play's full rating histogram (402,333 ratings) as the population baseline, so the review corpus is never mistaken for the player base.
What I'd need from you
- Retention curves by lifecycle stage × spend tier — the single most important table.
- New-user funnel, step by step: install → tutorial → first pull → first faction join.
- ARPDAU and payer conversion by segment, to replace the benchmark assumptions in the model.
- Gacha drop rates and pity mechanics, and whether odds have changed over time.
- What happened around the late-2025 team-mode change — see Act 4.
What 2,945 players actually said
One finding reframes everything else, and it replicated independently on both platforms.
The word “fun” appears at the same rate in 1–2★ reviews (19.2%) as in 4–5★ reviews (20.3%) — a skew of −1.1 points. For comparison, “love” skews −15.0 and “money” skews +26.7. Players who rate this game one star still call it fun.
That is not what a dying game looks like. A dying game loses the loop — players stop finding it fun and say so. Here the affection survives the one-star rating intact, and what replaces it is a vocabulary of money. “Money” is the single most-used word in the entire corpus — 898 uses, ahead of “fun” at 699.
Every word players used, sized by frequency
Colour shows which rating the word belongs to: red words are far more common in 1–2★ reviews, blue in 4–5★, grey in both about equally. Hover for exact counts.
n = 2,945 reviews · stop-words and game-name words removed · sized by √frequency
The words that separate a 1-star from a 5-star
Percentage-point difference in how often each word appears. Mention money and you are 26.7 points more likely to be leaving 1–2★. This is the most robust view here — because it compares rates within each group, sampling bias doesn't distort it.
Two more views: the true rating distribution, and theme sentiment
The true rating distribution — all 402,333 Play ratings
Not a sample: the whole population. The J-curve of a mature live game — a large satisfied majority with a hard negative tail. 70% five-star, 8% one-star.
Google Play, Aug 2026 · 32,790 of 402,333 raters wrote text (8.1%)
Theme reach vs. sentiment
Bar = share of reviews touching the theme. Number = average rating of those reviews. Sorted worst first: trust and odds is the most damaging theme in the corpus, at 2.06★.
Five pain points, in players' own words
Each one pairs the measured signal with a verbatim review and, where I have it, a screenshot from my own first session.
The store arrives before the game does
Money lexicon · 36% of reviews · avg 2.41★
At League 2 with a three-superstar roster, the store's featured shelf holds exactly one offer — $99.99 — and the memberships tab already lists four subscriptions. Value is asked for before it is delivered.
“The game's great at getting you hooked and wanting more but the store runs more like a casino… Game's great, store is trash.” 4★ · Daniel Fulda · Play · May 2025 · 52 upvotes
The social layer is locked exactly when it's needed
Faction/social · 7% of reviews · avg 2.70★
Faction unlocks at League 5, Showdown at League 4 — the entire activation window passes without them. Meanwhile global chat is open on day one and is wall-to-wall recruitment spam and inter-faction hostility.
“Chat is often terrible. The faction quests are too hard to achieve… It's not as fun as it used to be.” 4★ · Simon Heffernan · Play · Sep 2025 · 120 upvotes
Nobody explains the game
Complexity · 3% of reviews · avg 2.37★ · under-reported by design
Four upgrade systems, six gem stats, Trainer, Link & Gear and injury timers are all exposed before a first meaningful win. Day-one churners don't write reviews, so this theme is small in the corpus and large in reality.
“No idea what I'm doing, very brief tutorials that vanish instantly… tells you you're low health but no idea what you're supposed to do about it.” 1★ · Joseph Adler · Play · Apr 2026
Progress stops without a wallet
Progress/grind · 14% of reviews · avg 2.42★
Fuse Up costs 9,000 coins and 80 shards. A three-superstar roster is gated behind 59-minute injury timers, and the fail-state offers a 100-cash rescue. Events lock behind characters you don't have.
“So many events are locked because you don't have certain characters… The game is really fun besides that.” 2★ · Chris White · Play · Feb 2025 · 231 upvotes
The odds are invisible, and trust has gone
Trust/odds · 9% of reviews · avg 2.06★ — worst of any theme
The lowest-sentiment theme in the entire corpus. Players report duplicate-heavy "guaranteed" pulls and no published probabilities, and some have organised spending bans in response.
“They claim the pulls are random but a few of us got the exact pulls… Scopely also doesn't tell you the odds.” 1★ · WhyAds · App Store · Mar 2024
Everyone wants the same fix
The most common constructive suggestion in the corpus
Players aren't asking for the economy to be dismantled. They're asking for a version of the game that a non-spender can play alongside one that a spender can dominate.
“They really need to have two or three versions of this game. They have it so geared towards whales & big spenders that the average person cannot compete.” 1★ · Scottbodin · App Store · Oct 2025
See it — the first ten minutes, in screenshots
The first ten minutes, as I experienced them
Shown the party you can't joinLeague 2 · “No Faction” · Roster 3. “Unlock FACTION at League 5” and “Unlock Showdown at League 4”, with two panels greyed out.IMG_0832
Day-one global chatWall-to-wall recruitment with JOIN buttons — one ad demanding “discord mandatory” — interleaved with factions sniping at each other.IMG_0836
$99.99 is the whole storeThe League 2 shelf has one filled slot beside two empty ones. Below it, $14.99 on a 5h42m timer.IMG_0821
A third of the roster asleep“INJURED! Ends in 59m 34s” on one of three superstars. Fuse Up wants 9,000 coins and 80 shards.IMG_0843
The fail-state is a checkout“YOU'RE ABOUT TO LOSE!” → “RESCUE your Superstar for 100”, YES / NO.IMG_0847
Four subscriptions, no team yetKing of Kings, DX Army, DX Elite, Strap Up — all offered before the player has a roster.IMG_0822
What the store promises
The current App Store listing — what a new installer is sold about thirty seconds before they meet the League 2 store.
One thing I can't explain from outside
The two most-upvoted recent Android reviews, posted a day apart in December 2025, independently say the same thing: the faction/team layer was the best part of the game, and recent changes stripped it back toward solo play.
“The game has a great faction/team element but they are now taking away all the team modes… Lots of people have left the game” — 1★, 326 upvotes. And: “Recent changes have removed most team modes… Many long time players have quit” — 1★, 55 upvotes.
If that's accurate it changes the framing of idea 2 below, and the question I'd want answered first is what the retention data showed around that change.
One programme, and the mechanism behind each move
Three moves on one causal chain, all aimed at the same factor: the share of players who ever pay. None of them touches whale pricing or an existing drop rate, because the goal is more payers, not bigger ones — and those two get confused constantly.
Numerator · P0→P1
Earn the ask
No monetisation surface before a win. The first match cannot be lost, so the first thing the game ever asks you to buy stops being a rescue from a defeat. Systems reveal on a schedule; the store opens after a victory, not before a roster.
Mechanism. Today the first offer is a loss-tax shown to a player with three superstars and no win. Moving the ask behind the first win changes what is being sold from relief to momentum — the only framing an entry purchase converts on.
- Base-case value
- Moves
- Ever-pay +0.25pp
- Effort
- 6–8 wks · 1 pod
- Guardrail
- D30 ARPDAU non-inferior
Numerator · P0→P1→P2
Build the rung
Two rungs beneath the existing $14.99 starter pack, which becomes the top of the entry ladder rather than its floor. Both deterministic — the player sees exactly what they get — and built around a superstar they have already used. Rung 1 at the first wall, rung 2 within 72 hours of rung 1. Once per account, no contents shared with a whale SKU, and the six-hour countdown goes.
Mechanism. The chain depends on this move and it is still the smallest of the three by value, because a $14.99 rung already exists — so this deepens a ladder rather than creating one. Structurally load-bearing and financially third; both are true, and the sequencing argument is what justifies building it anyway.
- Base-case value
- Moves
- Ever-pay +0.35pp
- Effort
- 5–7 wks · 1 pod + economy
- Guardrail
- Top-decile ARPPU — kill metric
Numerator · P1→P2→P3
Make it safe to buy
Publish the pull probabilities. Convert duplicates to shards toward a star the player chooses. Scoped to the entry band first, where the rungs are deterministic anyway — so it touches no existing drop rate. Published odds ships ahead of the rest: it is largely configuration and legal review, and it needs to be live before the rungs are.
Mechanism. Trust is the lowest-sentiment theme in the corpus at 2.06★, and a first purchase is an act of trust. This is also the only move that lifts repeat purchase, which is where the model's compounding lives — and where its softest assumption sits.
- Base-case value
- Moves
- Ever-pay +0.20pp · repeat +7pp
- Effort
- 4–6 wks · economy
- Guardrail
- Gacha revenue per payer
Why in this order — each necessary, none sufficient. 02 without 01 puts a great offer in front of someone who hasn't won yet. 02 without 03 converts them once and never twice. 01 and 03 without 02 is a clean, honest, well-sequenced shop whose cheapest item is still $14.99.
Parked, and the condition each one is waiting on
- Rookie Crew — the social on-ramp before League 5. A D30 play, and a good one, but it retains players rather than converting them, so it builds the denominator. Ships once there is a payer base worth retaining, and remains gated on whether faction-join is a top-2 retention correlate.
- Mid-game catch-up track — blocked on faucet-vs-sink data, as before. Also a denominator play: it keeps people alive rather than converting them.
- Tentpole win-back — cheap, and the dormant pool is enormous. Sequenced after move 02, so returning players land on a real ladder instead of the shelf that lost them the first time.
- Deeper whale extraction — rejected outright. No headroom, and the corpus prices the goodwill cost. This programme adds payers; it does not enlarge the ones we have.
What the programme is worth
A payer-funnel model you can audit, not a forecast. Every input is editable below — change any of them and both tables recalculate. The ones I would replace with your telemetry in week one are marked in red, and there are more of them than I would like.
Open the model — every input editable
The model in one line
ARPDAU = (payers ÷ DAU) × ARPPU. The moves act on ever-pay conversion —
the share of installs that ever make a purchase — and on repeat rate. Added payers are
priced at the marginal converter's value, not the incumbent's, because adding a cheap rung
raises conversion and lowers average first-purchase value. A power-law retention curve
(k = ln(D1/D7) / ln 7, D1 35%, D7 16%) is still fitted, but only to split total DAU into
new-install and legacy, and to price move 01's non-payer retention side-effect.
One correction to the previous model, carried into this one. Its base_dau
of 37,395 against $800k a month only reconciles at $0.22 ARPDAU if that figure is
DAU attributable to new installs (90,000 × 12.5 user-days ÷ 30), not total DAU. Total DAU is
therefore ~121,000, including the 9.5-year legacy base. That was ignorable when the metric was
retention. It is load-bearing now, because ARPDAU is measured against total DAU.
Inputs — edit any value
Assumption. Third-party estimate I could not verify at source.
Assumption. Genre range is roughly $0.15–0.40.
Assumption. Sets the DAU denominator with ARPDAU.
%. Mid-core benchmark — this is data pull 1.
USD. The softest input in the model — data pull 2.
USD. Today's converters skew high because the ladder does.
%. Carries the compounding term — data pull 4.
| Move | Assumed lift | Effort | Added payers / mo | Conservative | Base case | Optimistic |
|---|---|---|---|---|---|---|
| Programme, all three | ||||||
Annualised incremental revenue. “Added payers” is the base-case monthly increase in installs that go on to make a purchase.
The comparison that decides the portfolio
| Portfolio | Δ revenue / yr | Δ DAU | Δ ARPDAU | What it moves |
|---|
Both rows scored the same way against the same baseline. Marginal retained users are assumed to monetise at half the average ARPDAU, on the grounds that the users a retention bet adds are marginal survivors of the activation window who convert to payer below average.
Read this before you read the numbers. Scored the previous model's own way — every added user monetising at the full average — the retention leads are worth — a year and move ARPDAU by —. Neutral, and neutral by construction: that model multiplies added user-days by average ARPDAU, so the ratio cannot move. There is no assumption under which a retention bet raises ARPDAU. That is not a modelling choice; it is what arithmetic does when you add to a denominator.
The programme's figure is robust to the same correction because its gain comes from the numerator rather than from added user-days. That robustness, not the headline dollar figure, is the argument — and the programme's headline figure is the smaller of the two, deliberately shown rather than hidden.
Where this model is softest, in order
- The marginal converter's value. Adding a cheap rung raises conversion and lowers average first-purchase value. The model prices an added payer at a third of an incumbent's — but the gap between those two figures is modelled, not measured, and it drives the headline more than anything else here.
- The repeat-rate lift. It is the compounding term, and it carries move 03 to the top of the per-move table on the softest evidence in the document.
- The baseline itself. Ever-pay at 2.0% is a genre benchmark, not your number. So is the daily payer rate the ARPPU readout is derived from. Every one of these is a week-one query.
The proposal
Four data pulls to replace my assumptions, then one programme shipped as experiments with pre-agreed decision rules and a kill criterion written before the work starts. Full specs follow in Act 08.
Week one — replace my assumptions
- The payer funnel by session index. Install → session 1…7 → first store impression → first purchase. Where does a first purchase happen today, how long does it take, and at what price? This is the query the whole case is asking for.
- Price-point distribution of first purchases, and the SKU ladder a new account can actually see. Is there anything under $10 in front of a new player? What is the modal first-purchase price? This confirms or kills move 02.
- D7 and D30 for payers vs non-payers, matched on session count. The matching is the point — unmatched, it only tells you that engaged players spend. Sizes the downstream retention line and tests the asymmetry the whole priority order rests on.
- Second-purchase rate and time-to-second, for first-time payers. Sizes the compounding half of the model, currently its softest input.
Down from six. The two dropped — faucet-vs-sink mapping and the dupe-rate history — belong to parked work, and pulling them now would be answering questions nobody is about to act on. Support-ticket volume against churn is folded into pull 4, because a support failure is a second-purchase failure.
The kill criterion, written before the work starts. If pull 2 shows a sub-$10 SKU already visible to new players, and pull 1 shows ever-pay conversion at or above the 2% benchmark, move 02 is wrong. The floor is not $14.99, the premise fails, and this collapses back to a retention case — in which event the parked list is the real portfolio and I would say so. Better written down now than discovered in week three. And plainly: $14.99 is not sub-$10, so this criterion technically survived the 8 August correction — I do not want the threshold doing work the evidence should be doing.
Then build, in this order
| Phase | What ships | Gate to pass first | Window |
|---|---|---|---|
| 0 | The four pulls above; re-run the Act 06 model on real inputs | — | Week 1–2 |
| 1 | Move 01 · Earn the ask, plus the published-odds half of move 03 | Funnel shows the first ask landing before the first win | Week 3–8 |
| 2 | Move 02 · Build the rung — the three-rung ladder | Pull 2 confirms nothing sits below $14.99 | Week 5–12 |
| 3 | Dupe-to-shard, then the parked list in whatever order the data supports | Top-decile ARPPU held flat through the move 02 test | Week 13+ |
Published odds ships in phase 1 rather than with the rest of move 03 because it is largely configuration and legal review — already table stakes in several markets — and it needs to be live before the rungs are, not after. Phases 1 and 2 overlap deliberately: they touch different surfaces and can be split-tested independently on the same install cohort.
What I'd want to be wrong about
- That nothing sits below $14.99. I have already been wrong here once — I read this shelf as a single $99.99 offer and missed the starter pack beneath it. That is corrected at the top of this page, and it is the reason to trust the pull rather than the screenshot: the case rests on one shelf, one account, League 2, one day, and storefronts are routinely geo- and cohort-tested.
- That payer retention is causal rather than selection. Payers retain better; payers were also always the more engaged players. Pull 3 matches on session count to separate the two, and the downstream retention line is deliberately the smallest of the three revenue components so the case does not lean on it.
- That a cheap rung doesn't cannibalise the whale ladder. The guardrails — once per account, time-boxed, no shared contents — are designed for exactly this, but designs fail. Top-decile ARPPU is a kill metric on move 02 from day one of the test, not a post-hoc check.
The specs
Three one-pagers, written to be handed to a pod. Each states the problem, the hypothesis, what ships, what doesn't, how it's measured, and what would make us kill it.
01 · Earn the ask — no monetisation before a win
Ships first. Defer every economy surface until the player has a roster and a win, and remove the cash-rescue prompt from the first session entirely. Primary metric ever-pay conversion within 7 days; guardrail D1/D7 retention and D30 ARPDAU, both non-inferior. Base case — a year. 6–8 weeks, one pod.
Full spec — problem, flow, requirements, experiment, risks
A new player lands on a three-superstar roster that already exposes a four-layer upgrade ladder, six gem stats, Trainer, Link & Gear, Boss Breakers and injury timers. Within the first ten minutes they are offered a $99.99 pack as the only item on the League 2 store shelf, four subscription passes, and a padlocked Premium rewards column at zero points. The first losable match ends in a “RESCUE your Superstar for 100 cash” prompt.
So the first commercial proposition this game ever makes is a loss-tax, offered to someone who has not yet won anything. Value is asked for before any has been delivered, and the only thing on the shelf is priced for a player who does not exist yet.
If the first monetisation surface a player sees arrives after a win and sells momentum rather than relief, then the share of installs that go on to make a purchase rises — because an entry purchase converts on aspiration, not on rescue. Retention improves as a side-effect rather than as the goal, and nothing about whale-facing content changes.
Primary: ever-pay conversion within 7 days of install.
Secondary: time-to-first-win, session-1 completion, share of new players who see a
store surface before their first win — target zero.
Guardrail: D1/D7 retention and D30 ARPDAU, both non-inferior.
The honest read is that move 01's value is mostly enabling: it is what makes move 02's rung land on someone with a reason to want it.
Not touching drop rates, pack pricing, gacha odds, or any existing player's experience. Not a tutorial rewrite for players past League 3. Not a new-art project — this reuses existing screens and changes what is shown when. Not building the entry offers themselves; that is move 02.
Screen 1 — the debut pull. Cold open on a named marquee superstar, granted outright. No gacha animation, no currency, no RNG. Copy frames it as a debut, not a reward. One CTA: Claim your star. Bottom-nav Store and Loot are hidden, not greyed.
Screen 2 — the first match. Match-3 plus one finisher. Boss Breakers, Link & Gear and the six gem-crit stats are suppressed. A single coach line explains the finisher bar. The match cannot be lost: if the player's health would hit zero, the opponent's next turn is scripted to miss. No rescue prompt exists in this session.
Screen 3 — the win, and one next action. Victory, XP, coins, and exactly one contextual CTA: Level up The Rock. Not six systems — one verb. Completing it fires the Level Up unlock and ends session 1 on a return hook.
Screen 4 — the ask, earned. The first store surface the player ever sees opens after the win, on a roster they now care about, and its first shelf item is move 02's rung 1. Until move 02 ships, this screen shows the existing shelf and the experiment measures the sequencing change alone.
Systems unlock against triggers, one card per session, queued if they collide:
| Trigger | Reveals | Why there |
|---|---|---|
| Session 1 | Match-3 + finisher | The core loop, taught with one guaranteed win |
| First win | Level Up · the store, and rung 1 | First upgrade verb, and the first ask — both after value has landed |
| 2nd match / 2★ | Enhance | Opens the colour-class ladder once one star exists |
| Day-2 return | 2nd superstar + team | Rewards the return; introduces team-building |
| League 3 | Fuse · daily missions | Depth once a habit is forming |
| League 4 | Showdown · global chat | Competitive modes when a roster exists |
| League 5+ | Evolve · Boss Breakers · Link & Gear · crit stats | End-game complexity for engaged players only |
- R1. Guaranteed marquee pull fires on 100% of first sessions. No RNG on the debut pull.
- R2. At most two systems visible in session 1; all others gated to the schedule above.
- R3. Store, Loot, offers and subscription surfaces are unreachable until the first win.
- R4. The debut match cannot be lost, and no cash-rescue prompt can be reached in session 1.
- R5. Every screen exposes exactly one primary CTA.
- R6. Zero reachable dead-ends. Unavailable actions are hidden, never shown-then-blocked.
- R7. A skip path exists for reinstalls and veterans — never trap an experienced player.
Interrupt or offline: resume at the last completed step, not from scratch. Reveal collision: one unlock card per session, queue the rest. Marquee already owned (reinstall with restored account): substitute the next star in the debut pool. Accessibility: coach tips dismissible, reduced-motion honoured. Skip abuse: skipping forfeits the guaranteed pull, so it can't be farmed.
debut_start · marquee_claimed · first_match_win
(with time_to_win) · first_store_impression (with
had_won_before — must be 100% true against control's mixed rate)
· first_purchase (with price, sku,
session_index) · rescue_prompt_shown —
must be zero in session 1 · dead_end_hit ·
session1_complete.
All validated in the analytics pipeline before the experiment opens. Note that
first_purchase with price and session index is also data pull 1 — if it isn't already
instrumented, that is the first thing to fix regardless of whether this ships.
Unit: new installs, 50/50 by device cohort. Control: current onboarding. Variant: the sequenced flow above.
Decision rule: ship if 7-day ever-pay conversion clears the pre-registered MDE at p<0.05 and both retention guardrails are non-inferior. If conversion rises but D30 ARPDAU regresses, we moved purchases later without adding any — hold and iterate. If retention rises but conversion doesn't, we built a retention feature; say so plainly, and re-rank it against the parked list rather than claiming the win.
Deferred monetisation costs early revenue. Some share of players do convert on a day-1 offer, and the rescue prompt does convert somebody. The guardrail catches it; if D30 conversion drops, we reintroduce the ask earlier in the variant rather than abandoning the build.
A scripted unlosable match may read as fake. Mitigation: the assist is invisible (the opponent misses), not a revive prompt. Watch qualitative feedback and session-2 return.
02 · Build the rung — two deterministic rungs below the $14.99 floor
Structurally load-bearing, financially the smallest of the three. Add two entry-band offers beneath the $14.99 starter pack a new player can already see, both deterministic and once-per-account. Primary metric ever-pay conversion; kill metric top-decile ARPPU. Base case — a year. 5–7 weeks, one pod plus economy.
Full spec — problem, the ladder, requirements, experiment, risks
The League 2 featured tab carries a $99.99 League Essentials pack in the headline slot and, beneath it, a 2-Star Gold Andre Starter Pack at $14.99 marked “One Time Offer” on a 05h 42m countdown. The memberships tab lists four subscriptions. So a first rung does exist — priced at $14.99 and sold against a six-hour clock.
Every conversion funnel needs a first rung the customer can reach without deciding they are now A Person Who Spends Money On This. $14.99 is three to five times the conventional entry price for this genre, and there is nothing beneath it. The corpus is consistent with that reading: “money” is the most-used word across 2,945 reviews and the most-reached complaint theme at 36%, skewing +26.7 points negative — the vocabulary of people who would transact, describing a transaction they find hostile.
Stated as a limit, and as a correction: an earlier version of this spec said the shelf held only the $99.99 pack. That was a misreading of my own screenshot. It rests on one shelf, one account, one day, and storefronts are routinely geo- and cohort-tested. Data pull 2 confirms or kills it before anything is built.
If a new player is offered a small, deterministic, roster-relevant purchase at the moment they hit their first wall, then ever-pay conversion rises — and if a second rung sits between it and the existing $14.99 pack, a meaningful share climb rather than stopping at one. Without lowering the whale ceiling, because the new rungs sit below the existing ladder and share no contents with it, and whales pay for relative advantage rather than absolute cost.
The modelled lift was cut from +0.55pp to +0.35pp once the $14.99 pack was identified: two rungs beneath an existing floor are worth less than a ladder built from nothing.
Primary: ever-pay conversion (share of installs making a first purchase within 30 days).
Secondary: second-purchase rate within 7 days of the first; rung-to-rung climb rate;
time-to-first-purchase.
Kill metric: top-decile ARPPU, monitored from day one of the test.
Guardrail: total revenue per install cohort, non-inferior.
Not changing any existing SKU's price, contents or availability. Not a sale, a discount or a limited-time markdown on anything that already exists. Not a subscription. Not a battle pass. Not touching gacha odds — the rungs contain no gacha at all.
Prices for the two new rungs are the conventional entry ladder for the genre and are placeholders until pull 2 returns the modal first-purchase price. The structure is the spec; the numbers are an input. The $14.99 row is not a proposal — it is what is on the shelf today, shown so the handoff is legible.
| Rung | Trigger | Contents | Why it converts |
|---|---|---|---|
| 1 · ~$2.99 new | First wall — first match lost after the debut, or first energy gate | Deterministic: shards for a superstar already on the roster, plus enough coins for one visible upgrade | Solves the problem in front of them, at a price that isn't a decision |
| 2 · ~$6.99 new | Within 72h of rung 1, once only | A named superstar the player has already fought against, granted outright | The second purchase is the one nothing is built for today. This is the whole P1→P2 bet. |
| $14.99 already exists | Currently on a 05h 42m countdown from first sight | 2-Star Gold Andre Starter Pack — unchanged | Becomes the handoff at the top of the entry ladder instead of its floor. Drop the countdown: a six-hour clock on a stranger's first purchase is pressure, and move 03 exists to remove exactly that. |
- R1. Every rung is deterministic. The player sees the exact contents before purchase. No RNG anywhere in the entry band.
- R2. Every rung is once per account, ever. Not once per season, not once per event.
- R3. Rung contents are personalised to the player's actual roster — never a superstar they have no relationship with.
- R4. New rungs are visible for 72h so they clear the shelf rather than becoming permanent cheap inventory — but no countdown timer appears on a player's first-ever offer. The existing $14.99 pack loses its 05h 42m clock.
- R5. No rung shares contents with any SKU above $19.99. Enforced as a content-catalogue constraint, checked in CI, not by convention.
- R6. Rung 2 is not rendered to a player who has not bought rung 1. No teasing an offer that cannot be taken.
- R7. No rung may appear before the player's first win (move 01's R3 covers the surface; this covers the offer).
- R8. Top-decile ARPPU is instrumented and dashboarded before the first rung goes live.
Player never hits a wall: rung 1 falls back to a session-3 trigger rather than never firing. Player buys an existing large pack first: the ladder is suppressed — they are not the cohort. Reinstall with restored account: rungs already consumed stay consumed. Regional pricing: rungs are defined by store price tier, not USD, so the entry band is the local entry band. Refund on a rung: does not restore the offer; monitored, because a refund on rung 1 is the clearest possible second-purchase signal.
rung_shown and rung_purchased (per rung, with trigger reason and
time-since-install) · first_purchase (price, sku, session index) ·
second_purchase (with days_since_first) ·
rung_expired_unbought · arppu_top_decile as a daily series ·
refund_on_entry_sku.
Unit: new installs, 50/50, held for 30 days minimum — long enough to read the second purchase, not just the first. Control: current shelf. Variant: the three-rung ladder.
Decision rule: ship if ever-pay conversion clears the pre-registered MDE at p<0.05 and revenue per install cohort is non-inferior and top-decile ARPPU has not moved. Stop immediately if top-decile ARPPU declines beyond the pre-agreed threshold — that is cannibalisation and it is not worth any conversion gain.
The interesting failure mode to watch for: conversion up, revenue per cohort flat. That means the rung captured purchases that would have happened anyway at a higher price. It is a real outcome, it is not a win, and the ladder would need re-pricing rather than re-designing.
Cannibalisation of the existing ladder. The primary risk, addressed by R2, R4, R5 and the kill metric. Once-per-account and no-shared-contents are what make a cheap rung structurally unable to substitute for a large pack.
The premise is wrong and something already sits below $14.99. I have been wrong on this shelf once already — see the correction at the top of this page. Pull 2, before any build.
Anchoring damage. A $2.99 offer next to a $99.99 offer can make the latter look worse rather than the former look reachable. Mitigated by the rungs being time-boxed and contextual rather than permanently shelved beside the large packs.
03 · Make it safe to buy — published odds and dupe-to-shard
Cheapest of the three, and it scores highest. Publish pull probabilities; convert duplicates to shards toward a star the player chooses. Scoped to the entry band, so no existing drop rate changes. Primary metric second-purchase rate; guardrail gacha revenue per payer. Base case — a year. 4–6 weeks, economy — published odds ships inside phase 1.
Full spec — problem, scope, requirements, experiment, risks
Trust & odds is the lowest-rated theme in the entire corpus at 2.06★. A first purchase in a gacha economy is an act of trust, and right now it is placed blind. Separately, bugs & support reaches 20% of reviews at 2.42★ — the second-largest theme — with “support”, “scopely” and “customer” all in the ten most negative-skewing words, and “fix” used 252 times.
Under a retention thesis that was a side-finding. Under this one it is structural: you do not buy twice from a shop that lost your last order. What decides whether a first-time payer ever pays again is what happens when something goes wrong with the thing they bought.
If a player can see what they are buying before they buy it, and cannot be made worse off by a duplicate, then the second purchase becomes a decision rather than a gamble and repeat rate rises. Published odds also removes the single loudest constructive complaint players make, at close to zero engineering cost.
Primary: second-purchase rate within 30 days of a first purchase.
Secondary: ever-pay conversion; trust-theme sentiment in new reviews; refund rate on
entry SKUs.
Guardrail: gacha revenue per payer, non-inferior — publishing odds must not
suppress pulls.
Published odds — phase 1. Every pull surface displays its probability table, per rarity and per featured item. This describes the current rates. It does not change them. Largely configuration plus legal review; already regulatory table stakes in several markets.
Dupe-to-shard — phase 3. A duplicate converts into shards toward a star the player selects. Scoped initially to duplicates originating from entry-band and early-game sources, so the exposure is bounded and readable before it goes wide.
Out of scope: pity timers, changing any published rate, retroactive compensation for historical duplicates, and anything touching the whale-facing pull economy. Each is a defensible piece of work and each would make this move un-shippable in six weeks.
- R1. Odds are displayed on the pull surface itself, not buried in a settings menu or a support article.
- R2. Published values are generated from the live drop tables at build time. A hand-maintained odds table is worse than none — it will drift, and drift here is a trust catastrophe.
- R3. Entry-band rungs are labelled explicitly as containing no random element.
- R4. Dupe-to-shard conversion rate is published alongside the odds.
- R5. The player chooses the shard target; the game never chooses for them.
- R6. Legal sign-off on the disclosure wording per market before rollout, not after.
odds_panel_view (with surface and dwell_ms) ·
pull_after_odds_view · dupe_converted (with target star) ·
second_purchase · refund_on_entry_sku · gacha revenue per payer
as a daily series against control.
Published odds cannot be split-tested honestly — showing one cohort the probabilities and hiding them from another is not a test anyone should want to defend, and in some markets it is not legal. Ship it to everyone and read it as a before/after with a pre-registered analysis window and the seasonal baseline stated in advance.
Dupe-to-shard is split-testable and should be: 50/50 on new installs, 30-day hold, decision rule on second-purchase rate at p<0.05 with gacha revenue per payer non-inferior.
The odds are bad and publishing them makes it worse. The most honest risk on this page. If the true rates are indefensible, publishing is still the right call and the rates are the thing to fix — but that is a different, larger conversation and it should happen before the disclosure ships, not after a press cycle.
The repeat-rate assumption is the softest input in the model and it is what puts this move at the top of the per-move table. Data pull 4 sizes it before anyone commits.
Support capacity. Dupe-to-shard creates a new class of “where did my star go” ticket. Given that support is already the second-largest complaint theme, the tooling and the runbook ship with the feature, not after it.
Sources & method
Every number in this document, where it came from, and how much weight it can carry. Anything marked Assumption is mine, not measured.
The rule I applied. Anything I could fetch and re-fetch is stated flatly. Anything computed from that data is marked reproducible and the dataset is published. Anything I could not trace to a primary source is either labelled an assumption and made editable, or removed.
The revenue decline is the one I got wrong first time round — it read well and I could not source it. It's better to lose the headline than to defend a number I can't show you.
Corrections log. 8 Aug — I said the League 2 featured shelf held only a
$99.99 pack. It also carries a $14.99 2-Star Gold Andre Starter Pack on a 05h 42m
countdown. Found by re-reading my own screenshot; move 02's lift was cut from +0.55pp to
+0.35pp as a result. 8 Aug — the prototype captioned IMG_0829 as an
“Insufficient Currency” dead-end; it is the Manager's League rewards screen with a
subscription CTA and a locked Premium column. 8 Aug — the prototype claimed the match screen
IMG_0841 exposes Boss Breakers, Link & Gear and crit stats; those appear on the Trainer
panel IMG_0818. All three are fixed. Earlier — a claimed ~60% revenue decline was
withdrawn as unsourceable.
Every number in this document, and where it came from
| Number | Value | Source & method | Confidence |
|---|---|---|---|
| Installs | 50,000,000+ | Google Play store listing, minInstalls field, fetched via API 6 Aug 2026 | Verified |
| Play rating | 4.356 | Same call, score field | Verified |
| Play ratings count | 402,333 | Same call, ratings field | Verified |
| Rating histogram | 1★ 32,779 · 2★ 9,828 · 3★ 20,283 · 4★ 57,927 · 5★ 281,498 | Same call, histogram field. This is the full population, not a sample | Verified |
| Written reviews | 32,790 | Same call, reviews field. 8.1% write-rate is 32,790 ÷ 402,333 | Verified |
| Release date | 25 Jan 2017 | Same call, released field. Age computed to Aug 2026 | Verified |
| Current version | 0.900, updated 24 Jul 2026 | Same call, version / updated | Verified |
| iOS rating | 4.309 from 30,026 | iTunes Lookup API, id=1017432937, US storefront | Verified |
| Competitor store data | SuperCard 4.77 / 78,326 · E&P 4.69 / 101,353 · MPQ 4.58 / 27,508 | iTunes Lookup API, US storefront, id=775402833, 1117841866, 618349779, fetched 6 Aug 2026. Release dates and update dates from the same call. iOS-only, and rating counts are not installs | Verified |
| Competitor revenue | not used | Third-party estimates exist for every title in Act 01's comparison table, and are excluded for the same reason the decline figure was: the sources behind them require a login. The competitive argument is built only on store-API fields | Withheld |
| Review corpus | 2,945 unique | 474 from the iTunes customer-reviews RSS feed (10 pages across mostRecent + mostHelpful); 2,471 from Google Play via google-play-scraper (6 pages each across newest, rating, helpfulness). De-duplicated on review ID | Verified |
| “fun” skew | −1.1pp | Computed: share of 1–2★ reviews containing the token (19.2%, n=1,143) minus share of 4–5★ (20.3%, n=1,557). Tokeniser strips stop-words and game-name words | Reproducible |
| “money” count / skew | 898 uses · +26.7pp | Same method. Full derived dataset in data/review-analysis.json | Reproducible |
| Theme reach & sentiment | e.g. trust 9%, 2.06★ | Keyword families over the same corpus; a review can match several. Averages are of matching reviews | Reproducible, sample-biased |
| All verbatim quotes | — | Copied exactly from the two review feeds, with handle, rating, date and upvote count as returned by the API | Verified |
| Screenshots | 42 captured | Firsthand play session, League 1–2, July 2026. Captions re-checked against each image | Verified |
| League 2 store shelf | $99.99 headline pack; $14.99 starter pack, 05h 42m timer; 4 subscriptions | Firsthand session, IMG_0821, League 2 featured tab. Corrected 8 Aug — an earlier reading recorded only the $99.99 pack | Verified |
| Store screenshots | 6 shown | iTunes Lookup API, screenshotUrls | Verified |
| Monthly installs | ~90,000 | Assumption. Third-party estimates seen in search results; the Sensor Tower pages behind them require a login and I could not open a primary source | Assumption |
| Monthly revenue | ~$800,000 | Assumption. Same limitation. With ARPDAU it sets the DAU denominator the whole model is measured against | Assumption |
| ARPDAU | $0.22 | Assumption. Back-solved from the two rows above, and inside the typical mid-core range of roughly $0.15–0.40 | Assumption |
| Total DAU | ~121,000 | Derived. Monthly revenue ÷ (ARPDAU × 30). Splits into ~37,400 attributable to new installs (installs × lifetime user-days ÷ 30) and ~83,800 legacy base. An earlier draft used the 37,400 figure as if it were total DAU — corrected here, because ARPDAU is measured against the total | Reproducible |
| Ever-pay conversion | 2.0% of installs | Assumption. Mid-core RPG benchmark, not measured for this title. The single most important thing to replace — it is data pull 1 | Assumption |
| Marginal converter, 180d | $21 | Assumption. The 180-day value of a player who converts because an entry rung exists — deliberately set at a third of an incumbent converter's, since a cheap rung raises conversion and lowers average first-purchase value. Modelled, not measured, and it drives the headline more than any other input | Assumption |
| Existing converter, 180d | $62 | Assumption. Today's converters skew high because the visible ladder does. Used only for the repeat-rate term | Assumption |
| Second purchase in 30d | 35% | Assumption. Genre benchmark. Carries the compounding term, and is the softest input behind move 03's ranking — data pull 4 | Assumption |
| Daily payer rate | 3.0% of DAU | Assumption. Used only to derive the ARPPU readout ($7.33 per paying DAU-day). Nothing in the model's outputs depends on it | Assumption |
| D1 / D7 retention | 35% / 16% | Assumption. Match-3 RPG genre benchmarks. Used only to split DAU and to price move 01's non-payer retention side-effect — no longer the headline driver it was in the retention-first draft | Assumption |
| Per-move lifts | see Act 06 | Assumption. My judgement, expressed as conservative / base / optimistic ranges rather than point estimates. The combined base case is +0.80pp on ever-pay conversion — a +40% relative lift, cut from +50% on 8 Aug once the $14.99 starter pack was identified, since two rungs beneath an existing floor are worth less than a ladder built from nothing. The surviving premise is that nothing sits below $14.99, and data pull 2 tests exactly that | Assumption |
| Revenue decline | withdrawn | An earlier draft claimed a ~60% decline from a 2021 peak. Could not be traced to a primary source, so it has been removed rather than caveated | Withdrawn |
How to reproduce this
The derived datasets are in the repository: data/review-analysis.json holds the word
frequencies, divergence scores, theme reach and population histogram;
data/roi-model.json holds the payer-funnel inputs, the per-move outputs and the
portfolio comparison, and is regenerated by scripts/build-model.py. The collection and
analysis scripts are in scripts/. Both feeds are public and unauthenticated, so the
corpus can be re-pulled and the numbers re-derived independently.