TL;DR: An algo's backtest is a promise; its performance review is the audit. Review every strategy monthly on seven numbers — trades, expectancy, profit factor, maximum drawdown, slippage against the backtest, cost drag and time-in-market — and retire anything whose live expectancy has fallen below zero after costs. In India the review has three extra lines: STT and per-order charges that punish high-frequency logic, business-income tax treatment, and SEBI's 2025 retail algo framework.


Last updated: 8 September 2026 · 10 min read

By Pulkit Mangal — F&O trader since 2017, founder of TradeDiary. Traded across Zerodha, Kotak and Dhan, ran rule-based strategies through broker APIs from 2020; built TradeDiary after losing ₹14L to positions I never reviewed properly.


The backtest said 71% win rate and a 2.4 profit factor. Three months live, the account is down ₹63,000 and the algo is still running — because nobody defined what "not working" would look like before switching it on.

That is the most common way retail algo trading fails in India. Not a bad idea, not a bug, but the absence of a performance review: a fixed, boring, monthly comparison of what the strategy is actually doing against what it was supposed to do. Discretionary traders at least feel a losing streak. An algo loses quietly, at scale, with perfect discipline.

This post gives you a review framework you can run in an hour a month, the seven metrics to run it on, a worked example with real-looking numbers, and the India-specific lines — charges, tax, SEBI's rules — that a backtest never included.

💡 Short on time? The review is only as good as the trade ledger under it. Start a free TradeDiary account, import your broker's tradebook, tag each trade with its strategy, and the expectancy, profit factor and drawdown per strategy are computed for you — the review becomes reading, not arithmetic.

What "algo trading performance review" actually means

An algo trading performance review is a periodic, written comparison of a strategy's live results against two baselines: its backtest (what it was expected to do) and its own history (whether it is decaying). It answers one question — is this strategy still worth the capital it is using? — and produces one of three decisions: keep, cut size, or retire.

Three things make it different from just "looking at the P&L":

  1. It is per strategy, not per account. An account can be flat while one algo is quietly profitable and another is quietly bleeding. Every metric below is computed for each strategy on its own.
  2. It is net of the costs the backtest ignored. Most backtests assume free fills. In India, an algo that fires eight orders a day pays STT, exchange transaction charges, SEBI fees, stamp duty, GST and — on most brokers — a flat brokerage per executed order. On a small edge that is the whole edge.
  3. It has a kill rule written before the strategy goes live. "Retire if live expectancy is below zero after costs over 60 trades" is a review. "Let's see how it goes" is hope.

Three India-specific gotchas the review has to absorb:

Why it matters for Indian traders specifically

The retail algo wave arrived without a review habit

Broker APIs (Kite Connect, Dhan, Angel SmartAPI) and no-code platforms (Streak, Tradetron, AlgoTest) put strategy automation in front of hundreds of thousands of retail traders. What did not come with them was the institutional discipline of a monthly strategy review. SEBI's own studies of individual F&O traders — 89% of whom lost money in FY22 — describe a population that mostly does not measure.

SEBI's 2025 retail algo framework makes it formal

SEBI's February 2025 circular on retail participation in algorithmic trading (in force from August 2025) routes retail algos through the broker: strategies above an order-rate threshold must be registered through the broker with the exchange, API access is tied to a static IP, and the broker is responsible for the orders the algo sends. Practically, that means your algo now has an identity and a paper trail — and a review that ties each registered algo to its live results is no longer optional housekeeping.

Costs decide the winners

Two algos with identical backtests can have opposite live results purely because one trades three times a day and the other thirty. The review is where you find out which side of that line you are on, and the only place cost drag is visible per strategy rather than buried in the account total.

Backtests overfit; live data does not

A strategy tuned on three years of NIFTY data will look flawless on those three years. The performance review is the out-of-sample test you cannot fake — if live expectancy is a fraction of the backtest's, the backtest was fitted, not discovered.

Nobody else will do it

A fund has a risk desk. A retail algo trader has a Telegram group. The monthly review is the risk desk, and it has to be you.

The review framework: seven metrics per strategy

Compute each of these per strategy, on a rolling window of the last 60–100 trades (or the last quarter, whichever is longer), and write the number next to the backtest's equivalent.

Field Why it matters
Trade count Below ~60 trades, every other number is noise. State the count so the review knows how much to trust itself.
Expectancy per trade (₹, after costs) The average rupee result of one trade including all charges. Below zero after costs = the strategy is paying the exchange to exist.
Profit factor Gross profit ÷ gross loss. Under 1.2 live, the edge is too thin to survive a bad month; compare with the backtest's figure.
Maximum drawdown (₹ and %) Deepest peak-to-trough fall on this strategy's own equity curve. If live drawdown already exceeds the backtest's worst, the backtest was optimistic.
Slippage vs backtest Average fill price minus the backtest's assumed fill, per trade. Retail options algos routinely lose 0.5–2 points a leg here; on a scalping edge that is fatal.
Cost drag Total charges ÷ gross profit. Above 30–40%, the strategy trades too often for its edge.
Time in market / exposure Hours or days held per trade, and overnight exposure. Tells you what a gap will do to the drawdown figure.

Optional fields worth adding once the basics are habitual: win rate and average win/loss (to explain expectancy), Sharpe or Sortino on daily returns, worst single day, number of stop-loss hits versus target hits, and the count of manual interventions — every time you overrode the algo is a data point about trust.

How to run the review, step by step

  1. Export the tradebook, not the P&L statement. The tradebook has every fill with time and price; the P&L statement has totals. Zerodha Console, Dhan, Upstox and Angel all export a tradebook per segment. Import it into your journal so each fill is a row.
  2. Tag every trade with its strategy. A performance review is impossible on an untagged ledger. If the algo places orders with a tag (Kite Connect supports order tags), map it; if not, tag by instrument and time window once and keep the mapping.
  3. Reconcile fills against the algo's log. Missing fills, partial fills and rejected orders are the first sign of a slippage or latency problem. A 3% mismatch between what the algo thinks it did and what the broker executed is a review finding in itself.
  4. Compute the seven metrics per strategy on the rolling window. Put them in a two-column table: backtest | live.
  5. Apply the kill rules you wrote before going live. Typical ones: retire if expectancy after costs is negative over 60+ trades; halve size if live drawdown exceeds 1.5× the backtest's maximum; pause if cost drag exceeds 40%.
  6. Write three lines. What changed, why you think it changed (market regime, slippage, a code change on a given date), and what you decided. The written line is what stops next month's review from re-litigating this one.
  7. Diarise the next review for the same trade count or the same day next month. Reviews that depend on remembering do not happen.

A real example with numbers

Nikhil, 34, Pune, a backend engineer, runs three strategies through his broker's API on ₹8,00,000 of capital: a NIFTY option-selling algo that opens on expiry morning, a BANKNIFTY breakout algo, and a stock-futures mean-reversion algo he wrote last year. His first proper review, after 90 days:

Strategy Trades Win rate Net P&L (after costs) Expectancy / trade Max drawdown Cost drag
NIFTY expiry option selling 52 79% ₹1,18,400 ₹2,277 ₹41,200 (5.2%) 14%
BANKNIFTY breakout 138 44% −₹27,900 −₹202 ₹68,500 (8.6%) 47%
Stock futures mean-reversion 41 61% ₹19,300 ₹471 ₹33,000 (4.1%) 22%

The account is up about ₹1.1L, which is what Nikhil had been looking at. The review says something different. The breakout algo's backtest promised a 1.9 profit factor and ₹640 per trade; live, after ₹20 per order across 138 round trips plus STT and exchange charges, it is losing money with a cost drag of 47% — it is trading roughly twice as often as the backtest because it re-enters on every fresh breakout instead of once per session. The option-selling algo is carrying the account, but its drawdown of ₹41,200 came from a single expiry-day gap, and its 52-trade sample is thin. The mean-reversion algo is fine and boring.

Decisions written that evening: retire the breakout algo until the re-entry logic is fixed and re-backtested with realistic charges; keep the option seller at current size but add a hard stop sized for a 2× gap; leave mean-reversion alone and review again at 100 trades. Without the per-strategy review, the losing algo would have kept running inside a profitable account for months.

Common mistakes in algo performance reviews

  1. Reviewing the account instead of the strategy. A profitable account hides a losing algo. Fix: compute every metric per strategy tag.
  2. Trusting the backtest's fills. Backtests fill at the close or the mid; live fills at the ask, after latency. Fix: measure slippage per leg on live fills and feed it back into the backtest assumptions.
  3. Ignoring charges until the tax statement. Cost drag is invisible in a P&L that shows only net numbers. Fix: import a tradebook with charges, or estimate them per order in the journal, and track cost drag monthly.
  4. Reviewing on too few trades. Twenty trades can show a 70% win rate by luck. Fix: set a minimum trade count before any decision, and say so in the review.
  5. Moving the goalposts after going live. Deciding what "failure" means once the drawdown has already happened produces a rationalisation, not a review. Fix: write the kill rules before the first live order.
  6. Never counting manual overrides. Every time you pause, skip or hand-close an algo's trade, the strategy's live record stops meaning what you think it means. Fix: log interventions as their own tag and review that count too.

How TradeDiary helps

TradeDiary is built for exactly this ledger. Import the tradebook from any Indian broker, tag each trade with its strategy, and the journal analytics compute expectancy after charges, profit factor, win rate and maximum drawdown per strategy — with the equity curve and drawdown chart per tag — so the monthly review is a page you read rather than a spreadsheet you rebuild. Charges are itemised from the file or estimated per order from your broker's plan, which is what makes cost drag visible per strategy.

Start your review ledger free — import your tradebook and see each strategy's expectancy and drawdown in minutes.

You may also like: the trading expectancy formula — the maths behind the per-trade number above, maximum drawdown in trading — why the recovery is harder than the fall, what a trading edge actually is, R-multiple explained for sizing stops the review can trust, algo trading slippage — how far live fills drift from the backtest's, and walk forward analysis — how to test a strategy on data it never saw.

Frequently asked questions

How often should I review an algo trading strategy's performance? Monthly, or every 60–100 trades for a high-frequency strategy — whichever comes first. Less often and a losing algo runs too long; more often and you react to noise. The trade count matters more than the calendar: a strategy that fires twice a week needs a quarter to produce a reviewable sample.

What is a good expectancy for an algo strategy? Any positive expectancy after all charges is workable if the drawdown is survivable; the absolute number depends on capital and trade frequency. A useful check is expectancy divided by average charges per trade — if your edge is only two or three times your cost per trade, small changes in slippage or brokerage can wipe it out.

Why does my algo lose live when the backtest was profitable? The usual causes, in order: fills worse than the backtest assumed (slippage and latency), charges the backtest ignored, overfitting to the historical window, and the strategy trading more often live than in the test. Measuring slippage per leg and cost drag per strategy usually identifies which one within a single review.

Do SEBI's algo trading rules apply to a small retail trader? SEBI's February 2025 framework for retail algorithmic trading, in force from August 2025, applies to retail traders using broker APIs: strategies above the exchange-specified order-rate threshold must be registered through the broker, and API access is tied to a static IP. Check your broker's current implementation, since the operational details are set by the exchanges and brokers.

How are algo trading profits taxed in India? Intraday and F&O results are business income under Section 43(5), whether the orders were placed by hand or by code. That affects which ITR you file, how turnover is computed for the Section 44AB audit thresholds, and how losses carry forward. Keeping a trade-wise ledger with charges from the start makes the year-end numbers a report rather than a reconstruction.


This article is for educational purposes only and is not investment advice. Trading in equities and derivatives carries a risk of loss, including loss exceeding your capital in leveraged positions. Past performance and historical drawdown figures do not guarantee future results. Consult a SEBI-registered advisor before making investment decisions.