Why we don't publish a backtest

A backtest is a claim about a strategy. We measure components. Those are different things with different evidentiary bars.

Almost every tool in this category leads with a backtest. A curve that climbs from the bottom-left to the top-right, a CAGR figure, a max-drawdown number, and an implicit promise: run this and you'd have made that. StockLearn doesn't publish one. That's a deliberate choice, not an omission, and this piece explains the reasoning.

The short version: a backtest is a claim about a strategy, and we measure components, not strategies. Those are different activities with different evidentiary bars, and conflating them is how most retail-facing research goes wrong.

What a backtest actually requires

To backtest something you must first specify a complete trading system. At minimum that means an entry rule, an exit rule, a position size, a starting capital, a rule for what happens when several signals fire at once, and a rule for what happens when none do. Every one of those is a decision, and none of them come out of the data — you supply them.

That matters for two reasons. First, the resulting equity curve is a property of your choices at least as much as of the underlying signal. Change the exit from ten days to twenty and the curve changes shape entirely, using identical signal data. Second, and more importantly for us: once you publish a complete system with entries and exits, you are no longer describing the market. You are telling someone what to do with their money.

What we do instead

Our unit of measurement is narrower and duller. We take a single component — a volume spike, an EMA alignment, a MACD crossover, an open-interest state — and ask one question: over the next N sessions, did stocks showing this component behave differently from the market, and is the difference bigger than noise?

The output is an excess return over a fixed horizon and a t-statistic. Nothing else. No entry price, no stop, no sizing, no compounding. That number is a fact about the data. What you do with it is your decision, and we have no view on it.

The overfitting problem is worse than it looks

Suppose you test twenty variations of a rule and publish the best one. Under pure noise, roughly one of those twenty will clear a conventional significance threshold by luck alone. Publish that one and you have a beautiful curve describing nothing.

Backtests make this failure mode almost irresistible, because the parameter space is enormous. Entry threshold, exit horizon, filter conditions, universe, rebalance frequency — each is a dial, and each dial multiplies the number of variants you can quietly try before showing the winner. The published result carries no trace of the discarded ones.

Measuring a single component with a pre-specified horizon doesn't eliminate this, but it shrinks the surface enormously. There are far fewer dials to turn, and we publish the ones that came back negative or null alongside the ones that didn't. Several of our research guides exist specifically to report that a widely-believed effect didn't show up in our data.

Survivorship, look-ahead, and the quiet killers

Even an honestly-constructed backtest inherits every flaw in the underlying data. If your universe is today's listed stocks, you have silently excluded everything that delisted — and delisted companies are disproportionately the ones that went badly. If any input to your signal was revised after the fact, or if you used a closing price that wasn't knowable at the moment of entry, you have look-ahead bias and your curve is fiction.

These problems don't disappear when you measure components instead. But a component measurement makes them easier to see and state, because there's less machinery between the raw data and the number. A backtest buries them under an equity curve that looks authoritative regardless.

The regulatory line, stated plainly

ClusterMicro Technologies is not a SEBI-registered investment adviser or research analyst. In India, publishing specific buy and sell recommendations for compensation falls under those registrations. A backtested strategy with entries and exits, presented as something a reader could follow, sits uncomfortably close to that line at best.

We'd rather be clearly on the correct side of it than test how close we can get. Measuring what a condition did historically, and saying so, is description. Telling you to act on it is not something we do.

What this costs us

Honestly: it costs us persuasiveness. "This component showed a mean excess of −0.5% over three sessions with a t-statistic of −3.6" is a worse marketing line than a chart going up and to the right. People who want the second thing will not find it here.

What it buys is that our numbers mean what they say. When we report an effect, there's no hidden parameter search behind it, no assumed exit rule doing the work, and no implicit promise about your returns. That's a narrower product. We think it's a more honest one.

This guide is educational and explains how StockLearn interprets common technical indicators, using illustrative examples. It is not investment advice or a recommendation to buy or sell any security.