“95% accurate AI signal” is the most common marketing line in the Indian trading-tools space right now. It’s also, almost always, meaningless.
This guide breaks down what “accuracy” actually means for a trading signal, what numbers to demand from any AI product, and how to spot the tricks.
“Accuracy” is the wrong number
Win rate (the percent of trades that close in profit) is the only metric most apps quote. It’s also the easiest to fake.
Consider two strategies:
| Strategy | Win rate | Avg win | Avg loss | Expectancy |
|---|---|---|---|---|
| A | 90% | ₹100 | ₹2,000 | −₹110 per trade |
| B | 40% | ₹500 | ₹150 | +₹110 per trade |
Strategy A has a 90% win rate and is a guaranteed account-blower. Strategy B has a 40% win rate and prints money.
If an AI product leads with win rate alone, ask for the expectancy and average win vs average loss. If they can’t or won’t provide it, walk away.
The numbers that matter
Demand these for any AI signal product, on Indian data, with realistic costs:
- Expectancy per trade in rupees (or in R-multiples).
- Profit factor = sum of wins / sum of losses. > 1.3 is okay, > 1.6 is good.
- Max drawdown as a percentage of starting capital.
- Average win / average loss ratio (R-multiple).
- Calibration — if the AI says “high confidence”, does it really win more often?
- Out-of-sample performance — how did it do on data the model didn’t train on?
Any product that can’t show these has not done the work.
What “AI accuracy” usually hides
Common marketing tricks on Indian platforms:
- Cherry-picked screenshots. The 10 best trades of 2025 do not equal the strategy.
- No costs included. STT, brokerage, slippage destroy paper edges. See intraday taxation for the cost context.
- Survivorship bias. Backtest on Nifty 50 as it exists today, ignoring stocks that were removed.
- Lookahead bias. Using data not actually available at signal time.
- Overfitting. The model memorized history; it doesn’t generalize.
- Conditional accuracy. “95% accurate on trend days” — but who tells you it’s a trend day in real time?
If any of these aren’t ruled out explicitly, treat the accuracy claim as zero.
Calibration: the most ignored metric
Calibration asks: when the model says it’s 80% confident, does it actually win 80% of the time?
A well-calibrated model:
- Says 90% confidence → wins 90% of the time.
- Says 60% confidence → wins 60% of the time.
A poorly calibrated model says 90% confidence and wins 55% of the time. It looks confident but isn’t. Most marketed “AI signals” are uncalibrated — they conflate strong activation patterns with high probability.
For the broader trade-offs see AI vs traditional screener.
The “AI overlay” framing that’s honest
Honest AI products don’t claim to “predict prices”. They claim to:
- Rank stocks for further research.
- Summarize indicator stacks (see RSI, MACD, VWAP).
- Filter news into actionable signal.
- Compress 60-page research reports.
That framing is verifiable. “Tomorrow’s winners” framing is unverifiable.
Backtest realism — the test
Ask a product for a backtest with all of these:
- ✅ Tradable universe (e.g., Nifty 500, not “any stock ever listed”).
- ✅ Lookback period of at least 5 years spanning multiple regimes (2020 COVID, 2022 bear, 2024 election, 2025 sideways).
- ✅ Realistic execution: STT, brokerage, slippage of at least 0.05% per side.
- ✅ Position sizing rule, not “1 share each”.
- ✅ Capacity test — does the signal still work at ₹1 crore notional?
- ✅ Out-of-sample window (last 12 months not used in training).
If any of these is missing, the backtest is decorative.
What AI does well (genuinely)
To be fair, AI does add value when used right:
- Compressing data. 100 stocks × 10 indicators → one ranked shortlist.
- Detecting unusual patterns. Volume profiles, sentiment shifts.
- Cross-asset context. What is the rupee doing? What are bond yields doing?
- Narrative generation. Producing a one-line “why this stock” reasoning per candidate.
The IntradayEdge dashboard uses AI exactly this way — as a narrative + ranking layer on top of deterministic indicators. It is research, not advice; see the disclaimer.
What to do as a buyer
Before paying for any AI trading product in India:
- Ask for a 5-year backtest with costs included.
- Ask for profit factor, expectancy, and max drawdown.
- Ask for a calibration plot (or reasoning).
- Take it on a free trial; track 30 trades on a small position.
- Compare against a free baseline: a Chartink screen for RSI(14) > 60 and MACD bullish. If the AI doesn’t beat that, you’re paying for branding.
Common questions
Is any AI signal genuinely 90%+ accurate intraday? Not at any meaningful expectancy. The only “90% accurate” strategies on retail data are ones where the average loss is 5–10× the average win.
Is past accuracy predictive? Loosely, yes — if the backtest is honest and the regime hasn’t changed. Regimes change in India often.
Should I avoid AI tools entirely? No. Just don’t outsource judgment. Use AI as a research accelerant; keep the trade decision.
For the foundational view, read AI stock analysis in India: what works in 2026.