Algo strategy rationale: 4 reasons a rule might work, and how to test
Most systematic rules rest on one of four reasons an edge could exist: trend persistence, mean reversion, a risk premium or a behavioral effect. Name the reason first, write it as a prediction that could be wrong, test it after costs on related markets and unseen data, and retest it on a schedule because reasons decay.
By Soren Huang · Published · updated · Editorial policy
Contents
- What rationale lies behind common algorithmic trading strategies?
- Why does a rule without a rationale usually fail out-of-sample?
- How do you turn a rationale into a testable rule?
- How do you check whether the logic holds across markets and out-of-sample?
- Why do rationales decay, and how do you notice in time?
- Frequently asked questions
- References
What rationale lies behind common algorithmic trading strategies?
An algo strategy rationale is the reason a trading rule should work at all, and almost every systematic rule rests on one of four: trend persistence, mean reversion, a risk premium or a behavioral effect. Naming which one a rule depends on, before the first backtest, is what separates a tested idea from a lucky fit.
Think of a rule as a claim about other people. Someone has to be on the other side of each trade, and the rationale says who that is and why they keep showing up. Searches for winning algorithmic trading strategies usually return lists of rule types. The lists rarely say why each one should work, and that missing sentence is the most important one. The four rationales below are the ones most published research and most trading desks fall back on.
- Trend persistence is the tendency of a move that has already started to continue for a while, because information spreads slowly and positions are built over days, not seconds. Moskowitz, Ooi and Pedersen (2012) found this pattern across many futures markets over horizons of one to twelve months. Breakout and moving-average rules rely on it; the momentum trading strategy on futures article tests one version.
- Mean reversion is the tendency of a stretched move to come part of the way back, because short-term pushes from impatient orders are later absorbed by patient ones. Fading a stretched move, or trading a return to an average, relies on it; see mean reversion on futures with a testable example.
- A risk premium is a payment that one group accepts for taking a risk another group wants to shed, such as a producer who sells futures forward to lock in a price, or an investor who holds the index through its drawdowns. Carry and roll rules, and a long bias in index futures, rely on it.
- A behavioral effect is a repeatable reaction of traders rather than a feature of the asset itself: chasing the open after an overnight gap, selling at the first loss, rebalancing at the end of a period. Session-open and end-of-period rules rely on it.
| Rationale | Trend persistence | Mean reversion | Risk premium | Behavioral effect |
|---|---|---|---|---|
| Source of the edge | Information and positions arrive slowly | Short-term pushes are absorbed later | One side pays to shed a risk | Traders react the same way each time |
| Who pays for it | Traders who join late or exit late | Traders who chase the stretched move | Hedgers and risk-averse holders | Traders acting on emotion or on a fixed schedule |
| Example rule type | Breakout of a 20-day high | Fade of a move far from its average | Carry or roll, long index bias | Session-open effect, end-of-month flow |
| What would break it | Faster information, crowding | A real regime change behind the move | The hedger leaves or the risk disappears | The reacting group changes or learns |
Why does a rule without a rationale usually fail out-of-sample?
A story is not proof, but it shrinks the number of rules you are willing to search, and the fewer candidates you compare on the same history, the less likely the best one is an accident. Out-of-sample data is the part of the price history that played no role in choosing the rule, and it is where accidents are exposed.
Data mining is searching many rules on one price history and keeping the one that scored best, whether or not there is a reason for it. The problem is not the search itself. It is that the best score among many tries carries a bonus that comes from luck, and the bonus grows with the number of tries. Bailey, Borwein, López de Prado and Zhu (2014) show that with enough variants, the best backtest looks excellent even when every variant is pure noise. Harvey and Liu (2015) argue that any result must be discounted for the number of tests behind it. The mechanics are covered in why the best backtest is usually inflated.
A rationale is a filter that runs before the search. If the story is "moves persist for days", you do not test a rule that enters on a 3-bar reversal of a 5-minute chart, because the story does not allow it. Every rule the story excludes is one fewer candidate competing for the title of best backtest, so the same winning score means more. Picture two funnels: a wide one that accepts any rule that fits the data, and a narrow one that accepts only rules the story allows. One winner comes out of each. The winner of the narrow funnel beat fewer rivals, so less of its score is a luck bonus.
There is a second benefit. A rule with a rationale fails in a way you can read. If a breakout rule stops working, you can ask whether moves still persist in that market, and look at the data to answer. A rule found by searching alone fails silently, because there was never a reason to check.
How do you turn a rationale into a testable rule?
Write the rationale as a prediction that could be wrong, then as an entry, exit and cost setting, so the backtest can contradict the story instead of only confirming it. A rationale that cannot be contradicted is a slogan, not a hypothesis.
- State the rationale in one sentence. "In this market, moves persist for several days because positions are built slowly."
- Turn it into a prediction with a number in it. "After a close at a 20-day high, the next 10 days are more often up than down, and the average move after costs is positive." The prediction names a horizon and a direction, so a flat or negative result refutes it.
- Write the rule. Entry: buy at the next bar's open after a 20-day high close. Exit: after 10 bars, or at a stop placed below the recent range. Both numbers come from the prediction, not from a search.
- Set costs and the session before running. Commission, slippage per side and next-bar execution, as in commission, slippage and the average trade. A rationale that only survives with zero costs is not an edge you can trade.
- Write down what result would contradict the prediction. For example: an after-cost average trade below zero, or fewer up than down outcomes over the horizon. Decide this before you look.
The last step is the one most traders skip. When the result contradicts the prediction, the honest move is to reject the story, not to try 15-day and 25-day highs until one works. Changing the numbers until the backtest agrees turns the test back into a search. When you set up the search in strategy generation, the parameter ranges and entry conditions you allow are exactly where the rationale is written down, so a range that only contains values the story permits keeps the search inside the narrow funnel.
Hypothetical example
Rationale: moves in an index future persist for days. Prediction: after a 20-day high close, the next 10 days are up more often than down, and the average trade after costs is positive. Rule: enter at the next open, exit after 10 bars or at a stop below the 10-day low. Costs: $6 per round trip and one tick of slippage per side. Pre-written contradiction: average trade after costs below $0, or fewer than half of outcomes positive. Result over 300 trades: 52 percent positive, average trade $38 after costs. The prediction is not contradicted. Had the result been 44 percent and −$12, the story would be rejected for that market, not retuned.
How do you check whether the logic holds across markets and out-of-sample?
A rule built on a real rationale should behave in the same direction on related markets, on neighbouring settings and on unseen data, and a result that appears in only one market or one window is evidence against the story, not for the rule. The direction matters more than the size, because the in-sample result always carries a selection bonus that the other tests do not.
Think of a doctor confirming a diagnosis. One test result can be noise; several independent tests pointing the same way are hard to explain by chance. The same holds for a rationale. If moves really persist in index futures, a 20-day breakout should show a positive after-cost average trade on ES, NQ and YM, not only on the one where it was found, and it should keep its sign in the out-of-sample block that the search never saw. The split and the untouched final block are described in how to split in-sample and out-of-sample.
| Check | What a real rationale shows | What a data-mined rule shows |
|---|---|---|
| Related markets | Same sign on most of them, weaker on some | Strong on the search market, mixed signs elsewhere |
| Neighbouring settings | 15-, 20- and 25-day versions all similar | One setting works, its neighbours do not |
| Out-of-sample block | Positive, somewhat smaller than in-sample | Flat or negative from the split date |
| Final untouched block | Same sign, tested once | Sign flips |
| Doubled slippage | Smaller but still positive average trade | Average trade falls below zero |
Two details make this check fair. First, "related" means the rationale applies, not that the market is popular: index futures share one kind of flow, gold shares less of it, so a weaker gold result is expected and a reversed one is informative. Second, the neighbouring-settings check is a test of the story, not of the optimizer. If 20 days works and 18 and 22 do not, the market is not telling you that moves persist for exactly 20 days; it is telling you the result is noise. The method for reading a parameter surface is in parameter stability: plateau or lucky spike. Write the pass count you require, such as four of the five checks above, before you run them, and the validation thresholds in strategy validation let you record that rule once and apply it to every candidate in the same way.
Why do rationales decay, and how do you notice in time?
Crowding, rule changes and shifting market structure can remove the reason an edge existed, so watch the after-cost average trade against its past spread and retest the rationale on a schedule rather than waiting for the drawdown to tell you. No rationale is permanent, because each one depends on a group of other traders who may change their behavior.
Sullivan, Timmermann and White (1999) tested a large set of technical rules on a long stock-index history and found that the rule which looked best in the earlier data did not keep its result in the later period. That is the general shape of decay: the pattern was real for a time, then the conditions behind it changed. Three causes are common, and each has an early sign.
| Cause of decay | Early sign to watch | Action |
|---|---|---|
| More traders using the same rule | After-cost average trade drifts toward the round-trip cost; fills get worse at the signal price | Compare the current average trade with its past range, not only with zero |
| An exchange or product change | Session hours, tick size, settlement or a new contract alter the session effect or the roll | Retest the rule on data after the change date only |
| A change in who trades the market | A related market still shows the sign while this one has lost it | Rerun the cross-market grid and see which cells changed |
The most useful single number is the after-cost average trade measured over the most recent block of trades, compared with the spread of single trades from the original test. A drop that is small relative to that spread is normal noise; a drop that moves the average close to the round-trip cost is the sign that crowding or structure has eaten the edge. The equity curve shows the same thing later, after the losses are already in the account.
Hypothetical example
Original test: average trade $38 after costs, single trades typically $320 away from that average, round-trip cost $6. A later block of 200 trades shows an average of $24. The uncertainty of a 200-trade average is about $320 ÷ 14 ≈ $23, so $24 is well within noise. A further block shows $4, with slippage per fill up from one tick to two. The average now sits below the round-trip cost and the fill cost has risen, which matches the crowding row above. The scheduled retest on the cross-market grid shows NQ still positive and ES flat, so the rationale is re-examined for ES before more capital goes to it.
A schedule beats a trigger. Set a fixed interval, such as every quarter or every few hundred trades, and at each interval rerun the full set of checks from the previous section on the newest data. The robustness checks that try to break a strategy are the same ones to repeat here, and the revalidation function in the strategy databank is built for exactly this kind of scheduled rerun. A rationale that still passes gets another interval. A rationale that fails gets replaced, with a new story and a fresh search, rather than a new setting on the old one.
The next step is a sheet with four lines for every strategy you run: the rationale in one sentence, the prediction with a number in it, the result that would contradict it, and the date of the next scheduled retest. Fill it in for the rule you trust most, and if the first line is blank, run the cross-market grid on it this week.
Frequently asked questions
What is the rationale behind a momentum strategy?
A momentum strategy relies on trend persistence: the tendency of a move that has started to continue for a while because information spreads slowly and large positions are built over days. Research on futures markets has found this pattern over horizons of roughly one to twelve months. The other side of the trade is traders who join late or exit late. The rationale is tested by predicting that returns after a breakout are more often positive than negative over a fixed horizon, after costs.
What is the rationale behind a mean reversion strategy?
A mean reversion strategy relies on short-term price pushes being absorbed later: impatient orders stretch the price, patient orders bring part of it back. The other side is traders who chase the stretched move. The test is a prediction that, after a move far from its average, the next few bars more often move back than continue. A real regime change behind the move is what breaks this rationale.
Can a strategy found by data mining still be worth trading?
It can, but only after it passes the same checks as a rule with a story: same sign on related markets, similar results on neighbouring settings, and a positive out-of-sample block it never saw. Those checks are harder to pass for a mined rule because the best of many tries carries a luck bonus. If it passes, try to write the rationale afterwards; if you cannot, treat the position size as a bet on an unexplained pattern.
How often should a trading rationale be retested?
On a fixed schedule rather than after a drawdown, for example every quarter or every few hundred trades. At each retest, compare the recent after-cost average trade with the spread of single trades from the original test and with the round-trip cost, and rerun the cross-market and out-of-sample checks on the newest data. A drop within the noise is normal; an average trade near the round-trip cost, or a sign that flips while a related market holds, calls for re-examining the story.
References
- Moskowitz, T. J., Ooi, Y. H., & Pedersen, L. H. (2012). Time Series Momentum. Journal of Financial Economics, 104(2), 228–250.
- Bailey, D. H., Borwein, J., López de Prado, M., & Zhu, Q. J. (2014). Pseudo-Mathematics and Financial Charlatanism: The Effects of Backtest Overfitting on Out-of-Sample Performance. Notices of the American Mathematical Society, 61(5), 458–471.
- Harvey, C. R., & Liu, Y. (2015). Backtesting. The Journal of Portfolio Management, 42(1), 13–28.
- Sullivan, R., Timmermann, A., & White, H. (1999). Data-Snooping, Technical Trading Rule Performance, and the Bootstrap. The Journal of Finance, 54(5), 1647–1691.

