Arbitraging an MMO economy with a homemade price tracker

The opportunity

As a student I played an MMO with a sprawling, player-run marketplace. People set their own prices, opened their own shops, even paid for extra listing visibility. A proper free market, and I was hooked: if you could see the prices move, you could buy low and sell high, with real in-game money on the table.

The whole opportunity is what happens when you line up one item across every shop selling it:

Reading a screen that didn’t want to be read

The game ran in its own desktop client. No browser, no API, nothing to intercept. I installed and configured at least five different OCR engines before admitting the fastest option was to write my own; I regret not starting there. Off-the-shelf OCR choked on the 9-pixel market font anyway. So I built my own reader: a fleet of game accounts saved a 77×11-pixel crop of each item’s best price, and a decoder thresholded the crop, sliced it into nine fixed digit cells, and read each digit through a hand-built decision tree that checks at most nine single pixels per cell. A “1” takes exactly one. The Java project’s real name is OdioGliOcr, “I hate OCRs”.

The machine around it

The reader was the middle of a pipeline that ran every night for about four and a half months, from autumn 2016 into February 2017:

• A dozen collector accounts (named Bread, Buff, Buy, Drop, Keep…) took turns searching the bazaar and screenshotting prices. The run logs read “ora raccolgo i prezzi”, “now I collect the prices”.
• Every crop became a row in a per-item CSV ledger: 33,292 readings across 193 items, from cheap consumables to the priciest gear in the game.
• A statistics pass computed an outlier-trimmed robust mean, plus 5/15/20-day windows and weekday and monthly means. A rule chain then issued Compra / Vendi verdicts, recorded which evidence fired, and routed the strong ones into instaBuy.txt and instaSell.txt, my morning shopping list.

It even drew charts for every item, straight out of 2017:

Rebuilt in the browser

In 2026 I dug the archive out and ported the tool’s logic to the browser, line for line. It isn’t a simulation: the page carries the real archived screenshots and the real price logs, and the port is verified against the archive. It re-decodes 1,677 stored crops into exactly the prices the tool logged in 2017, and a signal re-run “as of 10 Feb 2017, 18:00” reproduces the original’s final call for all 193 items, down to the flags each verdict fired and the means to the integer.

Launch the interactive demo ↗

Runs entirely in your browser. Watch the pixel decoder work through real screenshots probe by probe, then scrub months of real market history while the buy/sell engine re-issues its verdicts. The game stays unnamed.

What the returns looked like

I bought the lows and sold the highs, and at one point I was sitting on about three billion in in-game currency. That number is meaningless without the mechanics: the most expensive items in the game ran around fifteen billion, and there was a real-money market for that kind of currency. Trading it for real cash would have been a heavy terms-of-service violation, so I stayed clear. Part of me still regrets not daring.

The trades that paid best usually came from other people’s mistakes. That felt more sad than funny.

The other thing I got out of it was a close look at how an economy behaves when the underlying resources are effectively infinite, as they are in most games: prices anchor to habit more than to scarcity, which makes each deviation an unusually honest signal.

Why it ended

Four and a half months in, the fun ran out. I’d solved it, and a game loses its charm when you have full knowledge over the rules and how to win. If I did it again, I’d automate the buying and selling as well. I think I might already have; I don’t fully remember.

Notes

Nobody asked me to build any of this. I just don’t like leaving a system unread, games included. The nightly pipeline lived in a class called ProbablyNotAgileAndAlsoDeprecated.java.