Ask ten option sellers how they choose a backtesting tool and nine will talk about data, speed or the strategy builder. Very few will say what actually happens in their routine: the pricing model decides how many tests they run, and the number of tests decides which strategies they end up trading.
This is not a complaint about any vendor. Every pricing model has a reason. The point is that a cost model is also a behaviour model. Pay per run and you will think before you click. Pay a flat fee and you will click far more often. Pay nothing and you will click until the tool stops you. Each habit leaves a mark on your results, and the mark is mostly invisible in the final equity curve.
This article looks at the three common models in plain terms: credits, flat subscription and free tiers. It explains how each one bends research, shows the arithmetic of why the best of many tests is usually flattering, and ends with a simple ledger you can keep so that your tool's pricing does not choose your strategy for you. It names no vendor prices and makes no claims about any product. Those details change, so check them in the vendor's own documents.
A backtest has a cost even when the tool is free
Start with the idea that every backtest costs something. Money is the obvious part. The less obvious parts are waiting time, setup effort and the attention you spend reading the result. A tool that charges nothing per run may still charge you in queue time or in the number of runs allowed per day. A tool that charges per run may be cheap in time, because it returns results fast.
Economists call this the marginal cost: the price of one more run. It is a better lens than the monthly bill. Two tools can cost the same in a month and yet have very different marginal costs, because one lets you run twenty tests and the other lets you run two thousand.
What matters for research is that marginal cost sets your default behaviour. When one more run costs nothing, you run it. When one more run costs real money, you run it only if you expect it to teach you something. Both are reasonable. Both have a failure mode. The next section takes them in turn.
Three pricing models, three habits
Credit-based pricing: the rationing habit
Under a credit model you buy a pool of runs, or a pool of compute, and each backtest draws it down. The visible effect is that you slow down. You write the hypothesis before you run it. You choose the date range carefully. You think about whether a test is worth the credits.
That is good discipline. A trader who has to ask "what will I learn from this run?" is doing what a researcher should do. Credits also make the cost of curiosity visible, which stops the common habit of running a hundred small variations to see what sticks.
The failure mode is the opposite one. Rationing makes you protective of the first idea. If a run costs something, you are less willing to test the version that might disprove your favourite setup. You run the test you expect to pass. You skip the awkward neighbours: the same entry with a different exit time, the same strategy on the other index, the same rule on only the high-VIX days. Credits also push people to test over short date ranges to save cost, and a short range is where luck is loudest.
A second failure mode is quieter. When credits run out near the end of a billing period, people use the last of them on a rushed test and accept the answer. The decision to go live is then made by the calendar, not by the evidence.
How to use credits well: spend the first part of your allowance on the checks that could kill your idea, not confirm it. Run the long date range first. Run the cost-heavy version first. If the idea survives both, you have earned the right to optimise.
Flat subscriptions: the over-testing habit
A flat monthly fee turns the marginal cost of a run to near zero. This is the best model for exploration, because you can try a wild idea without guilt. It is also the model most likely to produce a beautiful backtest of nothing.
The mechanism is simple and it is worth seeing in numbers. Suppose you test a strategy variant, and the result is just noise: no real edge, only luck. Let us measure each result on a standard scale where zero means no edge and one means one standard deviation of luck. If you run one such test, the expected result is zero. If you run ten independent noise tests and keep the best, the expected best is about 1.5 on that scale. Run a hundred and the expected best is about 2.5. Run a thousand and it is about 3.2. These are standard results for the maximum of independent random draws, and they are not specific to trading.
Variants tested (pure noise) Expected best result, in units of luck
1 0
10 about 1.5
100 about 2.5
1,000 about 3.2
Read the table again with a flat subscription in mind. A trader who sweeps stop-loss from 10% to 50% in steps, entry time across 30 candidates, and two target settings has already run 5 x 30 x 2 = 300 variants, taking five stop values. The best of those 300 will look like a strong result even if the strategy has no edge at all. The more freedom the tool gives, the higher the best-of-many number climbs, and the less it means.
The honest answer is to count your trials and discount accordingly. The count is the single most useful piece of metadata in your research, and almost nobody records it. A flat fee does not make this problem. It removes the brake that would have made you notice it.
There is a second hidden cost on a flat plan: attention. Cheap runs invite long sessions. After the fortieth run of an evening, you are no longer reading results. You are scanning for green numbers. Set a limit on runs per session yourself, because the tool will not.
Free tiers: the constraint you did not choose
A free plan has a price too, and it is paid in limits. Typical limits in tools of every kind are a cap on date range, on number of saved strategies, on runs per day, on legs, or on features such as exports. They differ from vendor to vendor and change over time, so verify the current ones in the vendor's own documents instead of trusting a forum post.
The research problem with limits is that they pick your design for you. If a free tier shortens the allowed history, you test on a short window. If it caps the number of legs, you test a simpler structure than you would trade. If it blocks exports, you cannot run your own analysis on the trade list, so you rely on the summary the tool shows. Each limit shapes what you can learn, and none of them is shown in the equity curve.
Free tiers are still worth using. They are good for learning a platform and for a first pass at an idea. The mistake is to take a free-tier result, made on a shorter window and a simpler structure, and treat it as the answer for the real strategy. A useful rule: write down every limit that applied to a test next to the result. "Tested on last 6 months, 2 legs only, no costs" is a very different sentence from "tested".
Comparing models by behaviour, not by price
When people search for AlgoTest alternatives, the usual comparison is a table of features and a line for monthly price. A better question to put next to it is: what will this pricing make me do in an ordinary month? That is the question the table below tries to answer in general terms. It applies to any tool in any category, and the rows describe tendencies, not facts about any named product.
Model Marginal cost of one run Natural habit Main research risk Counter-habit
Credits Visible, every run Few, planned tests Protecting the first idea; short test windows Spend first credits on tests that could kill the idea
Flat subscription Near zero Many tests, wide sweeps Best-of-many bias; fatigue Keep a trial ledger; cap runs per session
Free tier Zero money, limits on scope Test whatever fits the limits Results from a narrowed design Record the limits beside each result
Per-seat or team plan Shared, hidden Others run tests on the same data Nobody owns the trial count One shared ledger across users
None of these rows says one model is better. A flat plan can suit a careful researcher who keeps a ledger. A credit plan can suit a careless one who runs the same broken test five times. What the table shows is that the pricing model sets the default, and the default is what most people follow.
The trial ledger: the habit that fixes all three
The fix is not a better tool. It is a file. Open a spreadsheet or a plain text note for each strategy idea, and record every backtest you run on it. Five columns are enough: date, what changed from the last run, date range, cost assumptions, and result. Add a sixth if you like: why you ran it.
This does three jobs. First, it gives you the number of trials. When you finally pick a winner, you can say "this was the best of 64 variants", and you can judge it against the table above. Second, it shows you the path. You can see whether you changed one thing at a time or wandered. Third, it works against the urge to forget bad runs. Most people remember the runs that looked good and quietly drop the rest.
With a ledger in hand, you can apply a simple discount rule. For every tenfold increase in the number of variants you tried, expect the best result to look better by roughly one unit of luck or a little less in the noise case. If your winning result is not clearly above what pure noise would give for your trial count, treat it as a hypothesis and not a finding. This is a rough guide, not a formula for truth, but it prevents the worst errors.
Thin samples make the problem worse
The number of trials is one side of the problem. The other is how much data each trial rests on. Consider a real example of how fast slicing eats a sample. EliteAlgo's desk study of 163 NIFTY and SENSEX weekly expiry sessions between January 2025 and July 2026 is a decent base for broad questions, with 83 NIFTY and 80 SENSEX days. Now split those sessions by India VIX band, as the study did for its premium measurements: 46 days under 12, 43 in the 12 to 14 band, 35 in 14 to 17, 24 in 17 to 20 and just 11 above 20.
The top band has 11 days. Any strategy statistic computed on 11 expiry sessions is fragile. Add a second slice, say separating NIFTY from SENSEX, and a cell has perhaps 5 or 6 days. Add a third (weekday of entry, or a particular entry time) and you may be studying two or three days. In a flat-fee tool, nothing stops you from slicing this way. The tool will happily print a profit factor for three trades.
The link to pricing is direct. Cheap runs encourage slicing, because each slice is one more run. Expensive runs discourage slicing but also discourage the wide, long, expensive run that would give you a decent sample. In both cases the number to watch is not the result. It is the sample size behind it. Put "n =" next to every result in the ledger.
What a sensible test budget looks like
A research budget can be set in plain proportions, whichever pricing model you use. One split that works for many individual traders is to divide your tests into three stages.
Stage 1: Exploration (about half of your runs)
Here you run wide and rough. Try different structures, entry times and exits. Use the longest date range your tool allows. Do not trust any single result. The goal is to find two or three ideas worth a closer look, and to drop the rest. Write down the number of variants you tried.
Stage 2: Stress (about a third)
Take the survivors and try to break them. Add realistic costs and a worse fill than you expect. Shift the entry time by a few minutes. Remove the best ten days and see what is left. Test the other index. Test only the high-VIX sessions. If the idea falls apart under small changes, it was a pattern in the noise.
Stage 3: Confirmation (the rest)
Choose one finalist and decide in advance what result would make you drop it. Then run it once on data it has not seen, or forward test it on paper. One run, one decision. If you did the earlier stages honestly, this stage needs very few runs, and that is why a credit plan is fine for it.
The split shows how pricing should map to stage. Flat plans and free plans are natural for Stage 1. Credit-style rationing is natural for Stages 2 and 3, where each run should have a reason. If your tool has one model only, you can imitate the other by self-imposed rules: a cap on exploration runs per week, or a rule that a confirmation run needs a written hypothesis first.
Costs inside the test matter as much as costs of the tool
One more link between price and research is easy to miss. The cost model inside your backtest decides whether the result is worth anything, and it is often a setting you can leave at zero. Brokerage, exchange charges, STT on option sales, GST and slippage all reduce what you keep. For short-option strategies that trade a small premium and exit at stop, slippage on the stop is usually the largest of these. Check the current charge schedule on your broker's site and on the exchange sites, such as nseindia.com, rather than using a number from memory.
Here is why this belongs in an article about pricing. A tool that makes it easy to leave costs at zero will, on average, produce more optimistic strategies. A tool that forces you to enter costs will produce fewer. The effect is similar to the trial-count effect: a quiet default tilts every result in the same direction. Whichever platform you use, set realistic costs before the first exploration run, not after you have found a winner. If you add costs at the end, you will find that the edge you chose is thinner than you thought, and you will already be attached to it.
As a rough guide to scale, consider an illustrative short straddle where you collect 100 points of premium on a lot of 65. That is Rs 6,500 of gross premium. If total friction (charges plus slippage on both legs, entry and exit) is an average of only 3 points, that is about Rs 195 per trade, or 3% of premium. If a stop-out adds 5 points of slippage on one leg, the friction rises to 8 points, or Rs 520, which is 8% of premium. Over a hundred trades, those are not rounding errors. They are the difference between a flat strategy and a losing one.
Reading backtests, yours and other people's
The same lens helps when you read a result someone else has published, whether on a forum, a vendor page or a social media thread. Ask five questions. How many variants were tried before this one was chosen? How many trades are behind the number? What date range was used, and was it chosen before or after seeing the results? Which costs were included? Which limits of the tool applied?
Most published backtests answer none of these. That does not mean they are wrong. It means you cannot tell. A result that comes with its trial count, sample size and cost settings is rare and worth more than a prettier curve that comes with none.
The same applies to your own work, six months later. When EliteAlgo publishes a study, such as the 163-session desk data used above, the useful parts are the sample size, the date range and the entry rule, and those are the parts to check first. A note that says "looked good in backtest" is useless. A ledger line that says "best of 37 variants, 163 sessions, costs on, 2 points stop slippage" lets you decide again with full information.
Choosing a tool by research behaviour
If you are picking a platform today, here is a short test list that has nothing to do with brand names.
First, find out the marginal cost of one more run in money, time and limits, and be honest about which habit it pushes you toward. Second, check whether costs and slippage are on by default or must be added. Third, check whether you can export the trade list so you can count samples and recompute statistics yourself. Fourth, check whether the tool keeps a history of your runs, because that makes a ledger much easier. Fifth, check how the tool treats dates: can you hold back a block of data and keep yourself from looking at it?
Then run the same simple strategy on two tools and compare the trade lists, not just the totals. Differences in fills, data and costs show up there. A tool that matches your live experience on a one-lot paper run is worth more than a tool with twice the features.
Frequently Asked Questions
Is a free backtesting tool good enough for option selling?
It can be good enough to learn the platform and to screen ideas. For a strategy you plan to trade with real money, check what the free tier limits: history length, legs, runs, exports. A result from a narrowed test should be repeated on a fuller set-up before you rely on it.
How many backtests should I run before going live?
There is no magic count. What matters is that you know the count, record it, and discount the winner for it. A handful of well-planned runs with a clear hypothesis beats hundreds of random sweeps. Always finish with a forward test on paper.
Why does my backtest look great but my live results are worse?
The common causes are best-of-many selection, costs set too low, unrealistic fills on stops, and a short or lucky date range. Start by checking your trial count and your cost settings. Then compare live fills with the fill assumed in the test.
Is credit-based pricing better than a monthly subscription?
Neither is better in general. Credits slow you down and can make you protect early ideas. Subscriptions speed you up and can invite over-testing. Pick the one that counters your own habit, and add the opposite rule yourself.
How do I avoid overfitting a strategy in a backtest?
Keep a trial ledger, test over a long range, change one thing at a time, stress the result with worse costs and shifted entry times, and confirm once on unseen data. A good introduction to the concept is available on investopedia.com.
Do backtest results include brokerage and taxes?
Only if the tool lets you enter them and you do. Many settings default to zero. Always check, and use the current charge schedule from your broker and the exchange.
Conclusion
The price you pay for a backtest decides how often you run one, and how often you run one decides how much of your result is luck. Credits ration you into protecting early ideas. Flat plans free you to over-test. Free plans narrow your design without telling you.
You cannot change the pricing, but you can change your habit. Keep a trial ledger. Put the sample size beside every result. Set realistic costs before the first run. Split your runs into explore, stress and confirm. These steps cost nothing on any platform, and they matter more than the choice of platform.
Algo trading carries a real risk of loss, and a good backtest does not remove it. EliteAlgo is analytical software and not an investment adviser, and nothing in this article is investment advice. The numbers above are either illustrative examples, standard statistical results or figures from a past-data study, and none of them predicts future returns.

Top comments (0)