Your First Systematic Strategy: What to Build (and What to Skip)

Skip the moving-average crossover. Your first systematic strategy should be an event study on a recurring, dated, disclosed event, and here is why.

Nearly every algorithmic trading tutorial starts the same way: compute two moving averages, buy when the fast one crosses the slow one, backtest on a decade of SPY, admire the equity curve. It is the "hello world" of the field, and it is a genuinely bad first project. Not because it loses money in a backtest (it often does not) but because of what it teaches you to do next.

Your first strategy has one real job, and it is not to make money. It is to install the habits you will use on every strategy after it: framing a claim, defining a population, measuring honestly, and attacking your own result. Judged by that standard, the crossover fails on every count, and a different kind of project, the event study, passes on every count. This post makes the case and gives you three concrete candidates you can build this week.

Why the moving-average crossover is a trap

Start with the obvious question the crossover cannot answer: why should it work? Who is on the other side of the trade, and what forces them to be there? A crossover is a pattern in past prices with no mechanism behind it, no participant who must act when the lines touch. When a backtest of it looks good, you have no way to distinguish a real effect from a lucky decade, because there is no reason for the effect to exist in the first place.

The deeper problem is what the crossover teaches you to do when the backtest looks bad: adjust the windows. Try 50 and 200, then 20 and 100, then add a filter, then another, until the curve looks right. That workflow has a name, curve fitting, and the crossover practically forces you into it because parameters are the only thing it has. Your first project ends up training the exact reflex that honest research exists to suppress. It also has no defined population: one instrument, one path of history, no way to ask whether the pattern held across hundreds of independent instances.

What a first strategy should actually teach you

Flip the requirements around and a checklist emerges. A good first project should have:

• A mechanism: a reason the effect exists, stated in terms of who is forced to act and why.
• A population: hundreds of independent events to measure across, not one price path.
• An unambiguous day zero: a specific date for each event, so the study design has no wiggle room.
• Few or no free parameters: nothing to tune, so nothing to curve-fit.
• A binary verdict: the claim survives measurement and attack, or it dies, and either outcome teaches you the process.

There is a class of market events that satisfies all five properties at once, and it is the class the free lesson on recurring, dated, disclosed events builds the whole discipline on.

The alternative: a dated, disclosed event study

The test is three words. An event class is worth studying when it is recurring (it happens again and again across many companies and years, so you study a population instead of an anecdote), dated (each instance happens at a specific, unambiguous moment, so day zero is never a judgement call), and disclosed (it leaves a paper trail, a filing or record, so a procedure can find every instance across the whole market).

Earnings dates, dilutive offering filings, lock-up expirations, ex-dividend dates, index changes: these all pass. A chart pattern fails all three parts, which is precisely why it makes a poor study. And the "everyone can see it coming" objection has a clean answer: the event is public, but the behavior it forces is lopsided. Index funds must buy the added name regardless of price. Option market makers must price earnings uncertainty before knowing the answer. Somebody in the room is trading on obligation, not opinion, and populations of forced behavior leave measurable footprints. The edge never lived in knowing the date. It lives in querying every event and measuring what the population did.

Three first studies you can actually run

Each of these is a complete first project: a claim, a point-in-time dataset, a measurement, and an attack. Pick the one closest to your interests.

• Earnings: implied vs realized moves. Claim: for large-cap names, the move implied by option prices before earnings is, on average, larger than the move realized after. The earnings move history dataset carries both numbers for each event, so the measurement is a few lines.
• Dilution drift. Claim: companies that file dilutive offerings drift lower, on average, in the weeks after the filing. The dilution events dataset gives you the dated, disclosed event stream; the attack step (borrow costs, tail losses) is where this one gets educational.
• Ex-dividend recovery. Claim: stocks recover their dividend drop within some horizon more often than not. The dividend capture dataset covers the calendar and the recovery outcomes, and the result will complicate a popular retail strategy in useful ways.

Whichever you pick, run it through the same four steps: hypothesis written first, point-in-time data, a handful of honest numbers, then a genuine attempt to kill the result by checking tails, regimes, sample size, and costs. Our walkthrough on running an event study in Python covers the code pattern, and the survivorship bias primer covers the most common way first studies quietly cheat.

What to skip, for now

A short list of common first projects that are better as fifth projects: machine learning models (you cannot debug a model's edge before you can measure a simple one), high-frequency anything (an infrastructure contest you have already lost), 24/7 crypto bots (execution complexity before research skill), and stacking technical indicators (the crossover trap with more parameters). None of these are permanently off-limits. They are just terrible teachers, because each one lets you build complexity faster than you can build honesty. If you are arriving from discretionary trading with existing convictions, codifying what you already believe is a better on-ramp than any of them.

Build it with your hands, graded

The Systematic Trading with Market Data course is built on exactly this progression: the free first module (no account needed) establishes why edges exist and which events are queryable, and the curriculum then runs the event-study loop across volatility, earnings, dilution, de-SPACs, dividends, and flows, in the browser, in real Python, against real market data, ending with a capstone where you design and defend an event study of your own. The recurring, dated, disclosed lesson is free right now, and the first lesson takes about ten minutes. Your first strategy is waiting in a filing feed, not in a pair of moving averages.