Backtesting runs your strategy's rules against historical market data to show how it would have performed. Done well it tells you whether an idea has ever worked. Done badly it produces a convincing number that means nothing, which is worse than no test at all — because a bad backtest gives you confidence you have not earned.
This guide covers what you need before you start, how to run one properly, which numbers actually matter, and the specific mistakes that make results look better than reality.
What you need before you start
- A strategy with rules precise enough to be executed mechanically
- An index to test on — NIFTY or SENSEX
- A date range long enough to include conditions you did not design for
- Entry and exit times, strike selection, and per-leg risk rules all defined
- A view on what result would make you reject the strategy
Decide your rejection criteria first
That last checklist item is the one almost nobody does, and it is the single most valuable habit in this whole process. Deciding in advance what would make you walk away stops you negotiating with the results afterwards.
Running one
1
Define the strategy
In the Option Strategy Builder: the index, each leg, strike selection, entry and exit times, and the risk rules.
2
Pick a date range
Longer is better, and it must contain more than one kind of market — trending, range-bound, and at least one volatility shock.
3
Run it
Each run uses one backtest credit. A bucket backtest uses one credit per active strategy inside it.
4
Read the report from the risk end
Start with max drawdown. Total profit is the least informative number on the page.
5
Nudge one parameter and re-run
A robust strategy degrades gently. One that collapses was fitted to this particular past.
Is ten a day enough?
Ten a day is enough for genuine iteration, including the fragility checks below. It is a cap, though, and if your process involves sweeping dozens of parameter combinations in an afternoon you will hit it — which, as it happens, is usually a sign the process needs changing more than the cap does.
What a backtest is actually simulating
It helps to be precise about what the engine does, because the gaps between that and reality are exactly where the surprises come from.
For each historical day in your range, the engine walks forward through the session. At your entry time it looks up the option chain as it stood, applies your strike selection rule, and records an entry at the price those strikes were trading at. Through the day it tracks each leg against your stop loss, target and trailing rules, and the combined position against your strategy-wide limits. At your exit time it closes whatever is open.
The assumptions baked in
What that process assumes:
- Every order fills. There is no queue and no rejection.
- It fills at the recorded price. No spread crossed, no partial fill.
- Margin is unlimited. The position is never refused for funds.
- You were there. The bot was running, the broker login was done, nothing broke.
None of those assumptions are unreasonable for a simulation. They are simply not true live, which is why the result is a ceiling rather than a forecast.
Which metrics actually matter
A backtest report gives you a lot of numbers. These are the ones that change decisions:
| Metric | What it tells you | What to watch for |
|---|---|---|
| Max drawdown | The worst peak-to-trough fall | Can you actually sit through this in rupees? |
| Drawdown duration | How long the worst stretch lasted | Months underwater breaks more traders than depth does |
| Win rate | Share of profitable trades | Meaningless alone — a 90% win rate with huge losers is a bad strategy |
| Reward to risk | Average win against average loss | Read it together with win rate, never apart |
| Expectancy | Average outcome per trade | The closest thing to a single honest summary |
| Max loss streak | Worst run of consecutive losses | Tells you the psychological load |
| Number of trades | Sample size | Under roughly 100 trades, treat conclusions as provisional |
Total profit is missing from that list on purpose. It is the number everyone looks at and the one most easily inflated by a lucky sample or a fortunate start date. Reading a backtest report properly goes through a full report line by line.
Choosing a test period
The instinct is to test on as much data as possible. That is roughly right, but the quality of the period matters more than its length.
What a good sample looks like
A period is useful if it contains:
- More than one market regime. A trending stretch and a range-bound stretch at minimum.
- At least one volatility shock. Events, results seasons, policy surprises — the days your strategy was not designed for.
- Several expiry cycles. Index option behaviour near expiry is different enough that a sample without them is incomplete.
- Gap opens. Overnight moves test whether your entry logic survives a price that is nowhere near yesterday's close.
The mistakes that make a backtest lie
Curve fitting
Adjusting parameters until the historical result looks good is not optimisation, it is memorisation. A strategy tuned to one specific past rarely survives a different future, and the tighter the fit the worse this gets.
The tell is sensitivity. Change the entry time by five minutes, or the stop loss by a few points, and re-run. If the result collapses, you have fitted the strategy to noise rather than to a real effect.
Selecting on the winner
If you test twenty variations and keep the best one, the winner's result includes a selection effect. Some of its apparent edge belongs to the search, not the market. The defence is to decide your parameters before you test rather than after.
Ignoring costs
Brokerage, exchange charges, STT, stamp duty and GST all apply per leg. A four-leg strategy pays them four times on entry and four times on exit. A strategy with a thin edge per trade can be profitable before costs and reliably unprofitable after them.
Treating the result as a forecast
A backtest is evidence about the past. It is not a prediction, and no honest platform will present it as one. Markets change, and a strategy that worked for three years can stop working without any warning that you would have seen in the data.
A worked sequence that avoids most of this
- Write the strategy down in plain English, including the exit rules.
- Decide the maximum drawdown you would accept, in rupees, at your intended size.
- Fix your parameters. Write them down so you cannot quietly revise them later.
- Run the backtest once over a long, varied period.
- Read max drawdown first. If it breaches the number from step two, stop — the strategy is too large for your account or simply not for you.
- Run the fragility check: nudge one parameter, re-run, confirm it degrades gently.
- If it survives both, paper trade it for at least two expiry cycles.
- Then go live at minimum size and compare the live period against the backtest for the same dates.
Steps two and five are what separate this from the usual process. Everything else is mechanics.
A worked example of a backtest that lies
Abstract warnings about curve fitting are easy to nod along to and hard to apply. Here is the concrete version.
A trader builds a short straddle on NIFTY with a 30-point per-leg stop loss and tests it over the last eight months. The result is good: steady profit, modest drawdown, a high win rate. Encouraged, they try tightening the stop to 25 points. The result improves. They try 22. Better again. They settle on 22 and deploy.
What went wrong
Live, the strategy stops out constantly and bleeds. What happened is that 22 points was not a better rule — it was the value that happened to sit just outside the noise of that particular eight-month sample. A slightly different period would have produced a different optimal number, and the live market is a different period.
What would have caught it
The diagnostic was available before deploying. Running 20, 22, 25, 28 and 30 and plotting the results would have shown a jagged line rather than a smooth one, and a jagged line means the parameter is fitting noise.
How much data is enough
There is no universal answer, but there is a usable way to think about it: count events, not days.
| What you want to have seen | Rough minimum |
|---|---|
| Expiry cycles | A dozen or more |
| Distinct market regimes | At least two |
| Volatility spikes | Two or three |
| Gap opens of meaningful size | Several |
| Total trades | Around a hundred |
A strategy that trades daily reaches a hundred trades in five months. One that trades only on expiry day needs considerably longer to reach the same confidence, which is worth knowing before you conclude it works.
After the backtest
A good backtest earns a strategy the right to be paper traded, not the right to be funded. Paper trading catches the problems a backtest structurally cannot see — and there are several.
If the strategy survives both, going live at minimum size is the next step. The gap between your paper results and your first live month is your real execution cost, and it is a number worth knowing before you scale.
Frequently asked questions
10 credits a day on Arthalab. A normal backtest uses one. A bucket backtest uses one credit per active strategy in the bucket, so a bucket of four costs four. Credits reset at midnight IST and do not carry over.
Enough to include more than one kind of market. A period that is entirely trending or entirely range-bound will flatter a strategy built for that condition. Quality of the sample matters more than its raw length.
There is no good number in isolation. A 40% win rate with large winners can beat a 90% win rate with occasional catastrophic losers. Read win rate and reward-to-risk together.
No. It means the idea worked on that data. Markets change, costs bite, and real fills are worse than simulated ones. A backtest is evidence, not a forecast.
Tuning parameters until historical results look good. The strategy ends up describing one specific past rather than any real market behaviour, and it usually fails going forward. Test for it by nudging a parameter and checking the result degrades gently.
NIFTY and SENSEX index options, on NSE and BSE.
Read your report carefully and check. Costs apply per leg, so a multi-leg strategy pays them several times on entry and again on exit — enough to turn a thin edge negative.
No. If a run fails to start, the credit is returned to you.
Start with a free 3-day trial
Build a strategy, backtest it and run it on paper — no broker, no IP and no money needed to try it.
