Backtest Lab · Nifty · 12 January 2023 to 26 May 2026
Does the AI beat the plain SuperTrend? A 3.5-year Nifty backtest
Not SEBI-registered. Educational content only. Every figure on this page is from a backtest on historical data, 12 January 2023 to 26 May 2026. It is not investment advice, a recommendation, or a forecast.
The short answer
StrikeATM backtested three SuperTrend indicators on 3.5 years of five-minute Nifty data, buying the at-the-money option on every flip, with real option prices and ₹100 of costs on every trade: TradingView's plain SuperTrend, LuxAlgo's SuperTrend AI, and AlgoAlpha's Machine Learning Adaptive SuperTrend. All three lost money, and both AI versions lost more than the plain one, because they flip more often and every flip pays costs. Selling options and holding to the next flip made money with the plain SuperTrend, but with a worst fall of almost half the capital, and a result a coin toss matched one time in five.
Key numbers from the StrikeATM backtest
- Buying on the flips, after costs: plain SuperTrend −₹67,772 (1,556 trades), LuxAlgo's AI −₹3,81,428 (2,490 trades), AlgoAlpha's ML version −₹1,59,200 (1,592 trades).
- Before costs, the plain SuperTrend made ₹87,828; AlgoAlpha's version made ₹0; LuxAlgo's lost ₹1,32,428.
- Signals from the option's own chart, and holding until the next flip, both lost too.
- Selling and holding with the plain SuperTrend: +70.8% in 3.5 years on an assumed ₹1,80,000 capital, with a 49.4% worst fall and 19 of 41 months lost. A coin toss did better 21% of the time.
- StrikeATM Score: 2.5–3 out of 10 for buying; 4.5 for selling and holding.
The two AI indicators we tested
Both are free and open on TradingView, and both are among its most popular scripts. Both call themselves AI or machine learning, and both are built on one of the most used indicators in Indian trading, the SuperTrend. We rebuilt each from its published code to test it; our versions are not shared.
- SuperTrend AI (Clustering), by LuxAlgo (licence CC BY-NC-SA 4.0). It runs nine SuperTrends side by side, from a tight setting to a loose one, scores how each did recently, sorts those scores into three groups with k-means clustering, and trades with the average setting of the best group.
- Machine Learning Adaptive SuperTrend, by AlgoAlpha (licence MPL 2.0). It sorts the market's recent volatility into high, medium and low with the same k-means method, and feeds that into an ordinary SuperTrend.
- The plain SuperTrend, TradingView's built-in (ATR 10, factor 3), as the benchmark.
Both are honest statistics. But k-means is a method from the 1950s: there is no language model here, nothing that reads the news. It is a clever way of choosing a setting.
How we tested it
The same rules for all three indicators:
- Chart: Nifty, five-minute candles.
- Signal: when the SuperTrend flips up, buy the at-the-money call of the current weekly expiry; when it flips down, buy the at-the-money put.
- Entry: at the next candle's open, because a flip is only known when its candle closes.
- Exit: when the indicator flips the other way, or at the close of the 15:15 candle. One lot each time, and ₹100 of costs on every trade, entry and exit together.
Before trusting any number, we checked our version of the LuxAlgo indicator against TradingView's own chart: every flip landed on the same candle.
Three real days
First, the kind of day these indicators are made for, then an ordinary one, on the plain SuperTrend and on the AI version.
The result: all three lost
Across 3.5 years of our backtest, buying on the flips, all three lost money after costs, and both AI versions lost more than the plain one.
| Indicator | Trades | Winners | After costs |
|---|---|---|---|
| Plain SuperTrend | 1,556 | 33% | −₹67,772 |
| SuperTrend AI (LuxAlgo) | 2,490 | 31% | −₹3,81,428 |
| ML Adaptive SuperTrend (AlgoAlpha) | 1,592 | 33% | −₹1,59,200 |
Where the money went
Before costs, the plain SuperTrend actually made money, ₹87,828; then its costs took ₹1,55,600. AlgoAlpha's version made ₹0 before costs: its machine learning added nothing. LuxAlgo's lost money even before costs, then paid costs on almost a thousand more trades. All three win only about one trade in three; the wins are bigger than the losses, about ₹3,182 against ₹1,639 for the plain SuperTrend, but not big enough.
| Indicator | Before costs | Costs | After costs |
|---|---|---|---|
| Plain SuperTrend | ₹87,828 | ₹1,55,600 | −₹67,772 |
| SuperTrend AI (LuxAlgo) | −₹1,32,428 | ₹2,49,000 | −₹3,81,428 |
| ML Adaptive SuperTrend (AlgoAlpha) | ₹0 | ₹1,59,200 | −₹1,59,200 |
On the option's own chart
Many traders put the SuperTrend on the option chart itself, not on Nifty. So we tested that too: each morning we took the at-the-money call and put, ran each indicator on that option's own candles, and bought whenever it flipped up. Same story.
| Indicator | Trades | After costs |
|---|---|---|
| Plain SuperTrend | 1,453 | −₹1,14,045 |
| SuperTrend AI (LuxAlgo) | 2,341 | −₹2,84,924 |
| ML Adaptive SuperTrend (AlgoAlpha) | 1,349 | −₹1,38,976 |
Maybe you exit too early?
The SuperTrend is a trend indicator, so why cut a trend every afternoon? We held every trade until the indicator flipped back, even across days, within the same weekly expiry. It got worse: the plain SuperTrend lost ₹2,11,378. Holding an option overnight means paying for time every night, and for a buyer, time is a cost.
| Indicator | Intraday | Held to the next flip |
|---|---|---|
| Plain SuperTrend | −₹67,772 | −₹2,11,378 |
| SuperTrend AI (LuxAlgo) | −₹3,81,428 | −₹3,38,018 |
| ML Adaptive SuperTrend (AlgoAlpha) | −₹1,59,200 | −₹2,95,014 |
Two things you will not see on the chart
LuxAlgo's version only calculates on the last 10,000 bars loaded on your chart. On the five-minute Nifty chart, that is only about the last 133 trading days; scroll back further and there is no line and no signals at all. Anyone who judged this indicator by scrolling back on TradingView has seen, at most, about six months of it.
AlgoAlpha's version can break. After a very large opening gap its volatility groups can come out empty, and the indicator turns to a sell with no line behind it (8 April 2026 above). In our data this happened on 15 days in 3.5 years. On 3 of them the sell alert fires but TradingView draws no arrow on the chart. The exact days depend on tiny differences in price data, so your chart may show it on different days.
What if you sell instead?
So far we have only bought options. Buyers pay for time; sellers collect it. So we flipped the trade around: when the SuperTrend flips up, sell the at-the-money put; when it flips down, sell the at-the-money call. Same entries, same exits, same costs. Intraday, it is still not much: the plain SuperTrend makes ₹16,939 in 3.5 years, and both AI versions still lose.
| Indicator | Selling, intraday |
|---|---|
| Plain SuperTrend | ₹16,939 |
| SuperTrend AI (LuxAlgo) | −₹1,79,356 |
| ML Adaptive SuperTrend (AlgoAlpha) | −₹41,160 |
Held to the next flip
Hold those sold options until the next flip, and the numbers change. To turn rupees into a percentage we need capital; selling needs margin, and margin changes from day to day. We assumed ₹1,80,000, as the capital needed across days over the 3.5 years. That is our assumption; the real figure would vary.
On that ₹1,80,000, the plain SuperTrend made ₹1,27,369: 70.8% in 3.5 years, about 21% a year or 1.73% a month on average. LuxAlgo's version made 43.2% and AlgoAlpha's 44.2% over the same period. Finally, something profitable. So, should you trade it?
| Indicator | Net | On assumed capital | Worst fall | Months lost |
|---|---|---|---|---|
| Plain SuperTrend | ₹1,27,369 | 70.8% | −49.4% | 19 of 41 |
| SuperTrend AI (LuxAlgo) | ₹77,699 | 43.2% | −96.7% | 20 of 41 |
| ML Adaptive SuperTrend (AlgoAlpha) | ₹79,613 | 44.2% | −67.5% | 21 of 41 |
The other side of that number
- The worst fall along the way was −₹88,926: 49.4% of the capital, almost half, gone before it came back. For LuxAlgo's version it was 96.7%.
- 19 of the 41 months lost money.
- The timing. For three full years it went almost nowhere: at the end of 2025 it was up just ₹6,346. Then, in the first five months of 2026, it made ₹1,21,024, and March 2026 alone made ₹90,809. Almost the whole result rests on a few weeks.
| Year | Net, sell and hold |
|---|---|
| 2023 | −₹40,809 |
| 2024 | ₹30,614 |
| 2025 | ₹16,541 |
| 2026 (Jan–May) | ₹1,21,024 |
The coin toss test
Is the SuperTrend choosing the right side, or does any option seller make money when the market does what it did? We kept every trade exactly as it was, same entry time, exit time and strike, but instead of the SuperTrend choosing whether to sell the put or the call, we tossed a coin, 200 times. Most of the coin tosses landed between −₹2.2 lakh and ₹2.5 lakh. The SuperTrend's result, ₹1.3 lakh, beat 79% of them.
That sounds good, but it is not good enough: roughly one random coin in five did better. To call something an edge we want it to beat at least 95% of random choices. This one does not.
The StrikeATM Score
The same ten-question check we use for every strategy in our course: costs, slippage, concentration, sample size, every year, tested or fitted. Buying on the flips scores 2.5 to 3 out of 10 for all three indicators. The version that made 70.8%, selling and holding with the plain SuperTrend, scores 4.5 out of 10.
| Variant | Plain SuperTrend | SuperTrend AI (LuxAlgo) | ML Adaptive SuperTrend (AlgoAlpha) |
|---|---|---|---|
| Buy, intraday, Nifty chart | 3 | 2.5 | 2.5 |
| Buy, intraday, option's chart | 2.5 | 2.5 | 2.5 |
| Buy, held to the next flip | 3 | 3 | 2.5 |
| Sell, intraday | 4.5 | 3 | 3 |
| Sell, held to the next flip | 4.5 | 4.5 | 4.5 |
The ten questions, for selling and holding with the plain SuperTrend:
| Question | Result | What we found |
|---|---|---|
| Look-ahead bias | Pass | Truncating the data at many points left every earlier indicator value identical; trades fill at the next candle's open. |
| Costs | Pass | Before costs, the average trade made ₹180; each trade then pays ₹100 in costs. |
| Slippage | Pass | The average winner is 42 times the slippage, and 2.4% of winners are smaller than it. |
| The curve | Fail | The worst fall was −₹88,926, 49.4% of the assumed ₹1,80,000 capital. |
| Trade concentration | Fail | Just 11 trades (0.69% of 1,588) made half of the entire profit. |
| Calendar spread | Fail | The best 5 months carry 86% of the profit. |
| Tested, or fitted? | Pass | Per trade: ₹5 in the first half of the data, ₹356 in the second. |
| Enough trades | Fail | t-statistic 0.94 on 1,588 trades: the profit is not distinguishable from luck. |
| Return per unit of wobble | Warning | Sharpe ratio 0.50. |
| Every year counts | Fail | 3 of 4 calendar years made money; the worst, 2023, lost ₹40,809. |
So, does the AI make the SuperTrend better?
On 3.5 years of Nifty data, in our backtest: no. Buying on the flips loses for all three; holding loses more; the AI versions lose the most, because they flip more and every flip costs money. Selling and holding made 70.8% with the plain SuperTrend, but with a fall of almost half the capital, almost half the months lost, nearly all the profit from one period, and a result a coin toss could match one time in five. A label that says AI does not make a strategy. A strategy is only as good as the questions it survives.
The same score, for the strategies we teach
For comparison, here are the option-selling strategies we build in the course, scored on the same card, on the same years, with costs of ₹160 a round trip.
| Strategy | Score | Average month (backtest) | Sharpe | Worst drawdown | Capital |
|---|---|---|---|---|---|
| Foundation I | 8.5 | 4.2% | 2.1 | −14% | ₹4,40,000 |
| Foundation II | 8.5 | 4.3% | 2.5 | −13% | ₹4,40,000 |
| Steady I | 9 | 6.0% | 3.4 | −13% | ₹4,40,000 |
| Steady II | 9 | 6.6% | 3.6 | −13% | ₹4,40,000 |
| Steady III | 8.5 | 7.0% | 3.4 | −22% | ₹4,40,000 |
| Steady IV | 8.5 | 6.7% | 3.1 | −22% | ₹4,40,000 |
| Shield I | 8 | 8.8% | 3.2 | −21% | ₹3,00,000 |
| Shield II | 8.5 | 8.2% | 3.4 | −18% | ₹3,20,000 |
| Shield III | 9 | 7.5% | 3.4 | −14% | ₹3,50,000 |
| Shield IV | 8.5 | 6.9% | 3.4 | −15% | ₹3,80,000 |
| Sprint | 8 | 13.2% | 2.9 | −53% | ₹2,20,000 |
None scores ten, and the course shows where each one loses its points. These are backtest figures on historical data, not a promise of future returns; each has its own drawdown and needs its own capital.
Frequently asked questions
Does LuxAlgo's SuperTrend AI work on Nifty?
Not in StrikeATM's backtest. Buying the at-the-money option on every flip on five-minute Nifty, 12 January 2023 to 26 May 2026, with ₹100 costs per trade, LuxAlgo's SuperTrend AI (Clustering) took 2,490 trades and lost ₹3,81,428 on one lot, more than the plain SuperTrend's ₹67,772. It lost money even before costs.
Does AlgoAlpha's Machine Learning Adaptive SuperTrend work on Nifty?
Not in StrikeATM's backtest. Buying on its flips, it made ₹0 before costs over 1,592 trades, so its machine learning added nothing, and lost ₹1,59,200 after costs.
Is the AI SuperTrend better than the normal SuperTrend?
In StrikeATM's Nifty backtest, no. Both AI versions lost more than TradingView's plain SuperTrend, mainly because they flip more often and every flip pays costs.
Does selling options on SuperTrend signals work?
In StrikeATM's backtest, selling and holding to the next flip with the plain SuperTrend made 70.8% over 3.5 years on an assumed ₹1,80,000 capital, but with a 49.4% worst fall, 19 of 41 months lost, almost all the profit in early 2026, and a coin toss beating it about one time in five. It scored 4.5 out of 10 on the StrikeATM Score.
Why does LuxAlgo's indicator show no signals on older Nifty data?
As StrikeATM found while backtesting it, it only calculates on the last 10,000 bars loaded on the chart, about 133 trading days of five-minute Nifty, so older candles have no line and no signals.
What data and costs did StrikeATM use?
StrikeATM used five-minute Nifty data and real at-the-money option prices of the current weekly expiry, 12 January 2023 to 26 May 2026, with ₹100 of costs on every trade (entry and exit together). Selling percentages use an assumed capital of ₹1,80,000.
Where can I see the full AI SuperTrend results?
In StrikeATM's Backtest Lab, at strikeatm.com/backtest-lab, with every table and chart, and as a video on StrikeATM's YouTube channel.
How we tested
- Data: five-minute Nifty candles (9:15 to 15:25) and real option prices of the current weekly expiry, 12 January 2023 to 26 May 2026.
- Indicators: TradingView's built-in SuperTrend (ATR 10, factor 3), LuxAlgo's SuperTrend AI (Clustering) and AlgoAlpha's Machine Learning Adaptive SuperTrend, at their default inputs, rebuilt from their published code; LuxAlgo's version checked flip by flip against TradingView's chart.
- Trades: at-the-money option, entry at the next candle's open after a flip, exit on the opposite flip or at the 15:15 close (or held to the next flip, where stated); one lot; ₹100 per trade.
- Selling percentages: on an assumed ₹1,80,000 capital; margin varies, so the real figure would vary.
- No look-ahead: truncating the data at many points left every earlier indicator value identical.
Is there an indicator you love that you want us to test the same way? Tell us in the comments on the video.
Not SEBI-registered. Educational content only. Backtest on historical data, 12 January 2023 to 26 May 2026. Indicators named are the property of their authors and are credited here; StrikeATM is not affiliated with LuxAlgo, AlgoAlpha or TradingView. Past results, real or simulated, do not guarantee future results. Nothing here is a recommendation to buy or sell any security.