Acumen

01 — The problem

Anyone can build a backtest that agrees with them.

A discretionary trader had a written intraday strategy he believed had an edge. The hard part was never running it. The hard part was running it in a way that could survive him not liking the result.

So the rules were frozen first, in a specification the code is never allowed to alter, and then replayed with no human judgement anywhere in the loop — every signal, every stock, no discretion, no skipped days. The strategy owns the rules; the engine only executes them.

It lost money in all eleven calendar years walked.

Trades taken

1,88,345

across 204 stocks and 2,428 trading days

Net result

−₹1.68 cr

against ₹1,00,000 of starting capital

Years negative

11 of 11

every calendar year in the span

Test cases

2,621

verified by running the suite, not by reading the badge

The whole result is one 3.07-point gap.

The strategy has a small positive edge before costs and a decisively negative one after. Every round trip costs ₹100 against ₹1,000 of risk — ten per cent of the money at stake before the market has moved at all — and the strategy takes 188,345 of them.

Reduced further, the ten-year loss is a single number missing its target. With this strategy’s own average winner and its own average loser, both after costs, it needed to win 34.60% of its trades to end level. It won 31.53%.

Not one good year to point at.

A losing strategy usually has a year that flatters it — a bull run, a volatile quarter, something to argue from. This one has none. Every bar is below the line, and the account never recovers the high it set on its fourth trading day.

The engineering is all in the audit trail.

Money is held in integer paise, never floating point, so the ledger reconciles exactly rather than approximately: gross 199,848,180 paise minus costs 1,883,450,000 equals net −1,683,601,820, and the report recounts it row by row before it prints. The volume-profile point of control is computed with exact fractions for the same reason.

Four pure deterministic functions carry the strategy. The same four drive a live Telegram screener that places no orders and never can — a structural guarantee, not a configuration flag. The screener was checked against the backtester over a fifteen-day sample and agreed candle for candle.

Source

49 modules

42,822 lines across the engine

Test cases

2,621

102 test files, 0 failures

Review reports

28

committed adversarial reviews under docs/reviews

Data lake

4.1 GB

4,95,312 symbol-days of minute candles

What it deliberately does not do

• It never places an order. The screener watches and sends a Telegram alert. A human always trades. That is enforced in the code path, not left to a setting.

• It does not correct for survivorship bias. The universe is today’s F&O list walked backwards, so a stock that left the index is absent and one that joined in 2024 is walked from 2016 anyway. That flatters any strategy trading liquid large caps. The report discloses it rather than quietly fixing it.

• It does not size positions realistically. The trader asked for the honest numbers with no limits, so every signal is taken. With no capital constraint the book peaked at 90 concurrent positions against ₹1 lakh of capital — which no real account could carry. Section 11 of the report measures the size of that assumption instead of hiding it.