Fifteen Years of StarCraft Research Before the Headlines

In 2003, Michael Buro, a computer science professor at the University of Alberta, argued that real-time strategy games were the next big test for artificial intelligence. Titles like WarCraft, StarCraft and Age of Empires forced a program to manage resources, guess at an opponent's strength, and reason about space and time all at once. Fifteen years later, the public conversation treats machines learning StarCraft as brand new.

It is not new. Academic work on the problem goes back to the turn of the century and has been a regular feature of the game AI community for well over fifteen years. DeepMind's partnership with Blizzard joined a long line of researchers rather than starting one.

Why does this belong on a finance blog? Investors face a similar problem: hidden information, limited resources, and many other players acting at once. The timeline also teaches patience, because a field can build real capability for years before anyone outside it pays attention.

There is a practical link as well. Large banks such as Bank of America (BAC) and Morgan Stanley (MS) have been hiring quantitative and machine learning staff for years, and their work draws on the same family of ideas: planning, search and decisions under uncertainty. We will start with the game itself, because the way it breaks computers explains a lot about how markets break forecasts.

Why a 1998 Strategy Game Still Beats Our Best Programs

StarCraft offers three factions, and each player begins with little. From there, they fight to control land and resources on a map covered by fog of war. Players scout, build structures and climb a technology tree that opens up stronger units, and every choice about what to build and when changes what the opponent can do next.

The early game is about finding the enemy while defending and constructing at the same time. Some players answer with a rush, a cheap early swarm meant to break the opponent before their economy matures (the Zerg rush gives the research its nickname). Others invest in economy and technology, accepting danger now in return for strength later. By the late game both sides are assembling army mixes, and the fighting turns chaotic.

Researchers call this a multi-agent control problem with imperfect information and an enormous state space. A program has to recognise which strategy it faces and switch plans on the fly. Chess and Go show every piece to both players, while StarCraft hides a large share of the board.

Computer scientists also measure difficulty formally. Classic NES and SNES titles such as Super Mario Bros., The Legend of Zelda and Donkey Kong Country have been proven NP-Hard, which already marks them as demanding. Real-time strategy games are considered at least NP-Hard and are predicted to be PSPACE-complete, or in the worst case EXPTIME, meaning that verifying a good solution could take exponential time.

That affects how we judge claims about progress. A system that masters a perfect-information board game has not thereby dealt with hidden information, and the same caution applies when someone says a trading model has "solved" markets. Markets hide even more than a StarCraft map does.

Open Platforms, Planning and Seven Years of Student Labor

Back in 2003, game companies were reluctant to open their engines or APIs. Research on StarCraft itself would have meant replicating, modifying or breaking the original game. So Buro led the creation of the Open Real Time Strategy platform, known as ORTS, a free and open source reduction of classic RTS games built for researchers and hobbyists.

ORTS grew slowly. Around 30 undergraduate and graduate students contributed over roughly seven years, and the platform gained scale, complexity and faithfulness to the commercial games along the way. That work is unglamorous, yet it gave a scattered community one shared testbed.

The research that followed included classical planning, an idea already used in commercial games such as F.E.A.R. and Empire: Total War. A planner describes the world as a state, lists actions with preconditions and effects, and searches for a sequence that reaches a goal. In StarCraft that becomes a build order: you need a certain structure before a certain unit, and enough minerals to pay for both.

Over time the original 1998 game gained competitions and benchmarks of its own, so bots written by different teams could be tested against each other. More recently, StarCraft II has drawn a surge of activity, including DeepMind's work with Blizzard.

The cause and effect here is plain. An open platform brought in students, students produced comparable results, and comparable results attracted attention. Shared measurement made cumulative progress possible, which is why the platform mattered more than any single paper.

Fog of War in a Portfolio: Where TIP, BAC, MS, EEM and UNG Fit

Investors never see the next inflation print or rate decision, so a portfolio is built under fog of war too. The defensive economy comes first, and for many people that means TIP, the iShares TIPS Bond ETF. It holds Treasury inflation-protected securities whose principal adjusts with the consumer price index, so it protects purchasing power.

The offensive units are BAC and MS. Both depend on rates and capital markets, but differently. BAC earns more from lending and deposits, so it tends to gain when short rates rise and the yield curve steepens, while MS leans on trading, investment banking and wealth management, which do better when markets are active and asset prices are climbing.

Then come the riskier pieces. EEM, the emerging markets ETF, carries currency and political risk, and a strong dollar tends to weigh on it. UNG tracks natural gas futures, and when later contracts cost more than near ones, the monthly roll erodes returns, so the fund can lose money while gas prices go sideways.

TIP has its own exposure. With a duration near 7 to 8 years, a one-point rise in real yields could cost roughly 7.5 percent of its value. The inflation protection also pays off only if inflation surprises on the upside.

Consider a hypothetical $100,000 portfolio: $35,000 in TIP, $20,000 in BAC, $15,000 in MS, $25,000 in EEM and $5,000 in UNG. In a stress case, EEM falls 20 percent (a loss of $5,000) and UNG falls 30 percent (a loss of $1,500), while TIP gains 2 percent ($700) and the banks stay flat. The net is a loss of $5,800, or 5.8 percent, against a $20,000 loss for someone who held only EEM.

Now suppose real yields rise one point. TIP drops about 7.5 percent, which is $2,625. If a steeper curve lifts the two banks 8 percent on $35,000, that adds $2,800, and the net is a gain of $175. These numbers are illustrations, not forecasts, but they show how the pieces offset one another.

Build Orders and Rebalancing Bands

Professional StarCraft players memorize build orders and then adjust them when scouting reveals something new. Investors can work the same way. Write the plan down, set target weights, and change course because of information, not mood.

A band system turns that into a rule. If EEM has a 25 percent target, you might rebalance only when it drifts above 30 or below 20. Checking quarterly keeps you from reacting to every headline while still catching real drift.

For entry, spreading purchases over time takes the guesswork out of timing. You could buy a new position in four monthly tranches of 25 percent each. For exits, trimming winners back to target is more reliable than trying to call a top.

UNG deserves different handling. Given the roll costs, it is a tactical position measured in weeks, so set an exit date when you buy it and keep it small.

Several practical snags show up. TIP's distributions are taxed as ordinary income, so many investors hold it in an IRA or 401(k). Selling BAC or MS in a taxable account can trigger capital gains, and buying back within 30 days after a loss runs into the wash sale rule.

Correlation is another trap. BAC and MS together make up 35 percent of the sample portfolio in one sector, and financial stocks tend to fall together. Treat the pair as a single bet when you size it.

What the StarCraft Timeline Suggests You Do Next

The research history carries two lessons that fit investing. Hard problems with hidden information yield slowly, through shared tools and years of patient work, and no one wins by rushing. A rush can break a weak opponent, but it leaves the economy thin if the opponent survives.

For a portfolio, that points to a written build order. Put target weights on paper for TIP, BAC, MS, EEM and UNG, or whatever your own holdings are. Decide in advance what each position is for: protection, rate exposure, growth or a short tactical bet.

Then set limits. Cap UNG at 5 percent, treat BAC and MS as one financial-sector position with a combined ceiling near 35 percent, and keep EEM inside a band such as 20 to 30 percent. Put TIP in a tax-advantaged account if you have room.

Pick two review dates a year and one rebalancing check each quarter. Write down why you hold each position, so that when markets turn chaotic you can see whether the reason has changed or only the price. If the reason holds, stay with the plan.

The numbers here are illustrations for thinking about structure and not personal advice. Your own time horizon, tax situation and tolerance for losses should set the actual weights.