The main families of algorithmic trading strategies, organized by the market condition each one depends on, and how to tell which is present.
Most lists of algorithmic trading strategies describe mechanics and stop there. The more useful organizing question is what each family assumes about the market. Every strategy encodes a bet about conditions, and it degrades when those conditions change.
This piece groups the main families by the condition each one depends on. For depth on the first two, start with mean reversion vs trend following.
A strategy is a rule plus an assumption. The rule is visible in the code. The assumption usually is not.
Trend following assumes moves persist. Mean reversion assumes they do not. Both are correct some of the time, and each loses money in the conditions the other needs.
This matters for a specific reason. A strategy that stops working has usually not broken. The market has moved into a state the strategy was never designed for, and nothing in the code will report that.
Reading the families through their assumptions makes the failure mode predictable rather than mysterious.
Trend following enters in the direction of an established move and holds while it continues. Moving average crossovers, breakout systems, and channel-based entries all sit here.
The assumption: directional moves persist long enough to cover the cost of entering late.
Where it fails: range-bound markets. Price crosses the average repeatedly without extending, and each crossing produces an entry that reverses. Losses arrive in clusters.
The characteristic profile is a low win rate with a large average win. Most trades lose a little, few trades pay for everything. That distribution is uncomfortable to run and is the reason many builders abandon it early.
Mean reversion enters against a move, expecting price to return to an average. Bollinger band fades, oscillator extremes, and pairs trading belong here.
The assumption: deviation from the average is temporary.
Where it fails: strong trends. The strategy adds against a move that keeps going. The win rate stays high while the losses get large, which is the most misleading equity curve in trading.
The profile inverts trend following. Many small wins, occasional severe losses. A short track record will flatter it, because the losing condition has not arrived yet.
Momentum ranks assets by recent relative performance and holds the leaders. Breakout enters when price clears a defined level.
The assumption: recent relative strength continues, and levels have meaning.
Where it fails: choppy conditions produce false breaks. Momentum also carries a rotation problem, since the ranking changes after the move rather than before it.
In crypto, momentum tends to be dominated by whether the whole market is moving together. When correlation across major pairs is high, ranking picks up beta rather than any asset-specific signal.
Arbitrage, market making, and funding rate capture aim to earn without a directional view.
The assumption: the spread or the funding differential persists, and execution is fast and cheap enough to harvest it.
Where it fails: volatility events. Market making accumulates inventory precisely when it is most costly to hold. Cross-venue arbitrage widens exactly when transfers are slowest.
These are the most infrastructure dependent strategies on the list. Edge lives in latency, fees, and execution quality rather than in the rule itself. For most independent builders they are the hardest family to run, not the safest.
Reading the families together produces a straightforward mapping.
The practical use of this mapping is as a gate rather than a switch. Running one strategy and disabling it in the wrong condition is more robust than rotating between families, because rotation multiplies the number of assumptions you have to be right about.
Market condition is measurable. ADX above 25 indicates a defined trend, with anything at 31 or below treated as weak. Moving average structure supplies direction. The RegimeLab dashboard shows that read across all tracked pairs on one screen, which is the same input a gate would consume.
A naive condition filter creates a new problem. Read the state on every candle and it flips constantly near the threshold, producing a stream of changes that are noise rather than information. A gate driven by that stream will disable and re-enable a strategy repeatedly within an hour.
Two fixes are available. Hysteresis uses separate thresholds for entering and leaving a state. Confirmation requires the same reading across several consecutive candles before accepting it.
We tested both against our own historical data. Hysteresis was rejected. Widening the band suppressed the noise, but it also discarded the moving average component of the classification, which meant genuine state changes were dropped along with the false ones. Confirmation across three consecutive reads removed the chatter without losing real events, at the cost of accepting the change later than it happened.
The general lesson transfers. When filtering noise, check what the filter removes rather than only whether the output looks calmer.
Knowing which family you run, and which condition it needs, is diagnostic rather than predictive.
It does not tell you the condition will hold. Every measure described here is computed from prices that already happened.
It does not create edge. A trend following system with no edge remains unprofitable in a trend.
It does not resolve position sizing, which fails independently of strategy choice and accounts for more damage than family selection does.
What it does is explain drawdowns you would otherwise read as randomness, and that is worth having before you change a working system.
The main families are trend following, mean reversion, momentum and breakout, and market neutral approaches such as arbitrage and market making. Each depends on a different market condition holding.
No family is reliably best. Profitability depends on whether the condition a strategy assumes is present, on execution costs, and on position sizing. The same rules produce different results in different market states.
Yes, though it multiplies the assumptions you need to be right about. Running one strategy with a condition gate that disables it is usually more robust than rotating between families.
Often what changed is the market state rather than the strategy. A rule that assumes trending conditions will underperform in a range without anything in the code reporting a fault.
ADX measures trend strength, with readings above 25 indicating a defined trend. Moving average structure supplies direction. Neither predicts what happens next, and both describe conditions that already occurred.
Algorithmic trading refers to automated execution of rules. Quantitative trading refers to strategies derived from statistical analysis. The terms overlap heavily and are often used interchangeably.
Not for most independent builders. Market neutral approaches depend heavily on latency, fees, and execution quality, and they carry concentrated risk during volatility events.