Algorithmic trading has a marketing problem in India. It gets sold as a way to stop thinking about the market, when it is really a way to stop improvising in the market. The strategy still has to be right. What automation buys you is that the strategy gets executed the same way at 09:16 on a quiet Tuesday and at 15:14 on the day the Nifty gaps 400 points. If you are still weighing whether to automate at all, automated vs manual trading sets out what automation does and does not change.
What counts as an algo
The regulatory definition is broader than the popular one. If software generates an order and sends it to the exchange without a human clicking "confirm" on that specific order, that is algorithmic trading. It does not matter whether the logic is a neural network or a single line that says "if the 20-EMA crosses above the 50-EMA, buy one lot".
This catches more setups than people expect. A TradingView alert wired to a webhook that fires an order is an algo. An Excel sheet with a broker plugin placing orders on a timer is an algo. A Python script that squares off your positions at 15:20 every day is an algo. All of them fall inside the same framework.
The four parts of any automated setup
Strip away the tooling and every automated trading system, from a one-file script to a co-located C++ engine, is the same four components in sequence. Being explicit about which one you are working on prevents a lot of wasted effort.
Data
Prices in, in a form your logic can consume. For intraday work on NSE this means a websocket feed from your broker, plus a decision about what you store and at what resolution. Historical data for research is a separate problem from live data for execution, and conflating them is the most common early mistake.
Signal
The rules that turn data into an intent: go long one lot of Bank Nifty futures. This is the part everyone spends their time on and it is rarely where the money is lost. A mediocre signal with good execution and risk control beats a good signal without them.
Execution
Turning intent into a filled order. Order type, limit price, how you handle a partial fill, what you do when the broker returns a rejection, what happens when the websocket drops for 90 seconds. This is unglamorous and it is where most live-vs-backtest divergence comes from.
Risk
The layer that decides not to trade. Maximum position size, maximum loss per day, maximum consecutive losses, a hard kill switch. Written before you go live, not after the first bad week.
The regulatory shape of it
SEBI’s February 2025 circular on safer participation of retail investors in algorithmic trading, and the exchange circulars that implement it, settled a question that had been ambiguous for years: retail traders may automate through broker APIs, and there is now a defined process for doing so. Full compliance has been required since 1 April 2026.
The practical consequences for someone running their own strategy on their own account:
- API access is tied to a static IP address that you register with your broker. A laptop on a home broadband connection with a dynamic IP will not do.
- Every algo order carries a unique identifier assigned through the exchange, so an order can be traced back to the strategy that produced it.
- An algo that sends 10 or more orders per second per exchange segment must be registered with the exchange. Below that threshold, registration is not required, but the orders still carry an algo tag.
- Your broker is the accountable party. If you use a third-party platform, the broker–provider relationship is formalised, and the broker answers to you.
Ten orders per second is a high bar for a discretionary-style systematic strategy and a low one for anything resembling market making. Most retail strategies — a handful of entries a day across a few instruments — sit nowhere near it. The rules in detail covers the registration process, the white-box/black-box distinction, and what changes if you sell or share a strategy.
Where beginners lose money
These are ordered by how early they bite, not by how much they cost. The expensive ones tend to arrive later.
Backtesting a strategy that cannot be executed
A backtest that enters at the close of the signal candle is assuming you knew the close before it happened. A backtest on a strike that traded 40 lots all day is assuming liquidity that was never there. Both produce equity curves that cannot be reproduced with real orders. Backtesting mistakes specific to Indian markets goes through the full list.
Ignoring the cost stack
Indian derivatives carry STT, exchange transaction charges, GST, stamp duty, SEBI turnover fees and brokerage. On options, the exchange transaction charge alone is levied on premium turnover, which makes high-frequency premium-selling strategies far more expensive than a naive per-trade brokerage estimate suggests. STT on futures and options both went up on 1 April 2026, which moved the break-even on short-holding-period strategies again.
No position sizing rule
A strategy with a genuine edge and no sizing discipline still goes to zero, because the sequence of losses that eventually arrives is longer than most people plan for. Size from the account, not from conviction.
Treating a disconnect as an edge case
Websockets drop. Broker APIs return 500s during the opening minutes. Your machine reboots. On a manual desk this is an inconvenience; on an automated one it is an open position with no supervisor. The correct assumption is that connectivity will fail during the worst possible five minutes, and the system should behave sensibly when it does.
| Assumption in a naive backtest | Live reality |
|---|---|
| Fill at the signal candle’s close | Fill somewhere after the candle closes, at a worse price |
| Zero slippage | Half a tick to several ticks, wider on illiquid strikes |
| Every order fills fully | Partial fills that leave an unbalanced multi-leg position |
| Costs are a flat per-trade fee | Six separate levies, some on notional, some on premium |
| Continuous data | Feed gaps, stale ticks, and reconnects |
| Margin is available when the signal fires | SPAN + exposure requirements move intraday |
A reasonable first three months
The ordering here matters more than the timeline. Compressing it is possible; skipping steps is what produces the horror stories.
Sequence, not checklist theatre
Pick a strategy you already trade manually
You need a prior for how it behaves. Automating an unfamiliar strategy means debugging the code and the edge at the same time.
Write the rules down until they are unambiguous
If two people reading your rules would place different orders, the rules are not finished.
Backtest with realistic costs and a pessimistic fill model
Assume you get the worse side of the spread. If the edge does not survive that, it does not exist.
Paper trade against the live feed
This catches the data and plumbing bugs that a historical backtest cannot, because it runs on the same feed and the same clock as production.
Go live at the smallest size the contract allows
One lot. The goal of the first month live is to find the operational bugs, not to make money.
Reconcile positions and P&L daily against the broker
Your system’s idea of your position and the broker’s must agree every single day. When they diverge, stop and find out why.
Only then, scale
And scale the size, not the number of strategies. Two half-understood strategies is worse than one understood strategy.
Where to go next
If you are still deciding whether this is worth the effort, read the cost article first — it is the fastest way to find out whether your strategy idea has any room to be profitable after fees. If you have already decided, the broker API comparison determines what you can actually build.
- SEBI’s retail algo rules, explained — registration thresholds, static IP, white-box vs black-box.
- Choosing a broker API in India — rate limits, websocket depth, order types.
- What F&O trading actually costs — the full levy stack with worked examples.
- Backtesting mistakes specific to Indian markets — survivorship, expiry rolls, look-ahead.