Designing Robust Investment Strategies

What "robust" means

A robust strategy is one whose behaviour does not depend on the exact period you tested, the exact parameters you chose, or the particular luck of a single run. Change the dates a little, nudge a threshold, and a robust strategy behaves roughly as it did before. A fragile one falls apart.

Notice that this is a different goal from finding the best-performing rules. The best-performing rules on any given stretch of history are, almost by construction, the ones most closely tailored to that stretch — which is exactly the property you do not want. Robustness is about durability, not about winning a historical contest.

This distinction is easy to agree with and difficult to act on, because the tools available to you measure historical performance and cannot measure durability directly. Everything below is about closing that gap.

Why a high CAGR isn't evidence

Here is the uncomfortable arithmetic at the centre of strategy design.

Any dataset — including several years of Indian equity prices — contains an enormous number of patterns. Most of them are accidents: coincidences produced by the particular sequence of events that happened to occur. If you search that dataset hard enough, testing rule after rule, you will certainly find rules that fit those accidents very well.

The result looks identical either way. Put the two side by side on everything a backtest can show you, and nothing separates them:

Only the last row separates them, and it is the one row a backtest of the design period cannot show you — because both strategies were measured on the very data that produced them.

So a strong historical return is not evidence of anything by itself. It is a starting observation that raises a question: is this the result of an idea, or the result of a search?

Two things push the answer towards "an idea": the logic was decided before the testing rather than discovered during it, and the result does not depend on any one setting being exactly right.

Overfitting and curve fitting

The two terms get used interchangeably and it is worth separating them, because one is the problem and the other is how you cause it.

Overfitting is the condition: a set of rules has learned the noise in your test data rather than anything general about markets, so it describes the past far better than it will describe the future. Curve fitting is the practice that produces it — adjusting rules until they match historical data closely.

Curve fitting is rarely deliberate. Almost nobody sets out to fit the past. It arrives one reasonable-looking change at a time, each justified by a slightly better result, and the strategy is thoroughly overfitted long before anyone decides to overfit it.

How curve fitting actually arrives

1. You have an ideaStated as a small number of rules
2. The result disappointsReasonable — most first attempts do
3. You add a conditionEach one sounds like a genuine improvement
4. The number improvesIt almost always will, which is the trap
5. Repeat until satisfiedThe strategy now describes the test period

Every individual step in that loop is defensible. That is precisely why the process is so hard to catch from the inside — at no point does it feel like anything other than careful work.

The tells are recognisable, and they are easier to see in someone else's strategy than your own:

The third one is the most useful diagnostic. A strategy expressing a real idea should tolerate small changes: if six-month momentum works, the neighbouring windows — three and nine months — should point the same way, even if not as strongly. If 6 works and both neighbours collapse, the strategy has learned something about the specific data rather than about momentum.

Design choices that tend to be durable

Robustness is mostly decided when the strategy is designed, not repaired afterwards. Creating Your First Investment Strategy covers how a beginner picks these values in the first place; what follows is the question to ask of each choice once you care whether it lasts.

Rebalancing deserves its own note, because it is where a design most often contradicts itself. The cadence should follow from the idea's timescale. A thesis about multi-month trends reviewed daily will trade constantly on movements the thesis does not care about; a fast-moving idea reviewed once a month will miss the changes it exists to capture. Neither is a wrong setting in isolation — both are mismatches with the rest of the design.

Underneath all five is one principle: prefer logic you could defend to a sceptical person who has not seen your results. If the only defence is the backtest, there is no defence.

Testing against conditions it wasn't built on

The single most useful thing you can do with a promising strategy is run it against a period you did not use while designing it.

This works because of what it risks. A strategy fitted to 2021–2023 has learned that period's particular character. Run it across a stretch with a different character — a sharp fall, a long flat drift, a rotation between market segments — and fitted strategies tend to deteriorate sharply while robust ones degrade gently.

NiveshBuddy's history begins in 2021, so "another period" means dividing what there is rather than reaching for a different decade. Design on one stretch — set the Backtest Period to Custom and choose, say, 2021 to 2023 — then run the unchanged rules over 2024 onwards. How far back you can reach also depends on your plan. A shorter history makes this split more valuable, not less: it is the only out-of-sample check you have.

What you are looking for is not a good result. It is consistency of behaviour — and each thing you check has a reassuring answer and a worrying one.

A strategy that returns less in an unfamiliar period but behaves the way its logic implies has passed something meaningful. One that returns well for reasons you cannot account for has not — it may simply have found a second coincidence.

Worth stating plainly: none of this makes a strategy reliable. Markets change in ways no historical test anticipates, and a robust design can still disappoint. What validation buys you is the removal of some strategies that were never going to work, which is a smaller claim and a real one.

Where Experiments fit

On plans that include it, Experiments run many variations of one strategy in a single pass — different ROC windows, holding counts, exit ranks and so on — and list every configuration ranked by the measure you choose. That makes them the most convenient sensitivity check in the product, and the most convenient way to overfit.

Used well, an experiment answers the question from the checklist above: does the result survive a nudge? Look at whether neighbouring configurations agree — the results page can expand close variants of each one for exactly this. A configuration whose neighbours do roughly as well is expressing an idea; one that stands alone at the top of a list of thousands has probably found a coincidence.

Used badly, an experiment is curve fitting at scale. Rank enough configurations on one historical window and the first-ranked one will always look excellent, because the ranking itself picked it for fitting that window. Treat the first place as a candidate to test on a period it was not ranked on, never as the answer.

Common misconceptions

• "A higher backtested return means a better strategy." It may mean a better strategy or a more thoroughly fitted one, and the number alone cannot say which.
• "More rules mean more sophistication." More rules mean a narrower eligible set and a closer fit to history. Sophistication is usually fewer rules that you can defend.
• "Robust means it won't have bad periods." Every equity strategy has bad periods. Robust means the bad periods are the kind its logic implies, not a surprise that reveals the logic was never there.
• "If it survives another period, it works." Surviving one more window is evidence, not proof. It removes some false candidates; it does not confirm the survivor.
• "I'll fix it by optimising the parameters." Searching harder over the same data — by hand or with an Experiment — is the mechanism that creates the problem, not the remedy for it. Use a search to check sensitivity, not to pick a winner.

Key takeaways