How EdgeTest works, and how I check it is right.

• Equity, daily, split and bonus adjusted (core set)

  Rows

  ~2.0M

  Range

  1991-01-02 to 2026-10-06

  Coverage and notes

  494 symbols traded on the latest day; about 340 have 15+ years

• Equity, daily, every listed stock

  Rows

  5.49M

  Range

  2016-01-01 to 2026-10-06

  Coverage and notes

  3,430 symbols traded on the latest day; 1,017 split events applied on read

• ETFs, REITs and InvITs, daily

  Rows

  254K

  Range

  2016-08-26 to 2026-10-06

  Coverage and notes

  376 instruments: 348 ETFs, 22 InvITs, 6 REITs

• F&O end of day

  Rows

  30.6M

  Range

  2016-01-01 to 2026-10-06

  Coverage and notes

  13.9M index option rows and 16.4M stock option rows (stock options from 2024)

• Implied volatility and Greeks

  Rows

  6.55M

  Range

  NIFTY from 2016-01-01

  Coverage and notes

  Black-Scholes, implied vol solved with Brent's method, computed in-house and recomputed nightly

• NIFTY options, 1-minute

  Rows

  ~58.2M

  Range

  2024-01-01 to 2025-09-30

  Coverage and notes

  435 sessions, 18,066 contracts, 93 expiries

• BANKNIFTY, FINNIFTY, MIDCPNIFTY options, 1-minute

  Rows

  ~64.8M

  Range

  2024-01 to 2025-09

  Coverage and notes

  33,146 contracts

• Equities, 1-minute

  Rows

  ~72.1M

  Range

  2024-01-01 to 2025-09-30

  Coverage and notes

  477 symbols

• Fundamentals, quarterly

  Rows

  41,146

  Range

  2017-03-31 to 2026-06-30

  Coverage and notes

  2,308 companies; 36 point-in-time metrics derived from these

• Corporate actions

  Rows

  23,011

  Range

  ex-dates 2016 to 2026

  Coverage and notes

  2,542 symbols, including announced future events

• India VIX

  Rows

  2,664

  Range

  2016-01-01 to 2026-10-06

  Coverage and notes

  Used for entry filters on options strategies

The three 1-minute sets add up to the 195M minute bars on the resume (58.2M + 64.8M + 72.1M).

• No look-ahead

  A condition seen at the close of day T fills at the open of day T+1.

• Chronology decides precedence

  A stop that fires during day T wins because it happened first. A pending signal exit fills at the next open before that bar's stop.

• Gap-through stops fill at the open

  If price gaps through a stop or target, the fill is the open, not the level you hoped for.

• A real cash-and-holdings portfolio

  It cannot overdraw, and the curve is a daily mark-to-market, so drawdown is bounded by construction. Signals dropped for want of cash, and positions opened smaller than asked, are counted and reported.

• Liquidity decides who gets a slot

  When more signals fire than there are slots, candidates are ranked most liquid first, by a 20-day trailing average of rupee turnover shifted one bar.

• Open positions are not dressed up

  Positions still open at the end close at the last price, are badged Still Open at End, and are charged no sell-side cost they never paid.

• Rules are a small language

  Any of 62 indicators, 7 operators (crosses_above, price_above_indicator and so on), left-to-right AND/OR, 15 tunable sub-parameters, and the same language on the exit side. A strategy with a stop but no way to close on profit is refused.

• Ranked baskets

  Hold the best N of a universe, re-decided weekly to quarterly, using 7 price and 19 fundamental factors with percentile ranking. 65 universes plus 22 industries.

• Up to 10 legs, each its own risk unit

  Buy or sell, call or put or a futures leg, strike rule, lots, and an expiry slot (current, next, or nearest monthly, which is what makes calendars, diagonals and covered calls on weeklies possible). Each leg can carry its own take-profit, stop multiple and trailing stop.

• Exits that compose

  Take profit, stop as a multiple of premium, max hold, close N days before expiry, overall rupee stop and target, square-off complete or partial, move survivors to break-even, lock profit with an optional ratchet, and re-entry with four trigger types.

• Settlement against the true cash close

  Hold-to-expiry settles at intrinsic value against the cash index close. The futures price was used earlier and was off by 55 points on average. Calendars and diagonals settle the expiring leg at intrinsic and sell the far leg at that day's market price.

• Margin gates entry

  A cycle the account cannot fund is not taken, and the count of skipped cycles is reported. Entry filters include an India VIX band and a 52-week ATM-weighted IV rank band.

• Undefined payoffs are withheld

  Payoff curves for two-expiry structures are not defined, so the report withholds them and says why, instead of drawing a plausible-looking line.

• Stressed days still trade

  If an exact strike did not trade on the entry day, the engine uses the nearest strike that did (within 2 steps) instead of dropping the cycle. This stopped months like March 2020 vanishing from backtests.

• Pin

  A run's window ends at the last completed session in the data, never at "today". A run at 11 am and one at 9 pm cannot span different data through no fault of the user. Historical end dates pass through untouched; the pin only ever clamps down, so it can never invent bars.

• Fingerprint

  A hash over the resolved inputs that determine the answer: engine version, cost-model version, data as-of date, resolved window, universe hash and config hash. It covers inputs only, never outputs, so the same fingerprint with different metrics is a bug alarm instead of noise.

• Replay

  Any report can be re-run exactly, and returns a verdict of reproduced, or exactly which input moved.

• Byte-identical regression

  Changes that should not move results are proven not to. In one release nine options shapes over three years were byte-identical to the previous engine, equity results were byte-identical on four shapes after the slippage change, and the minute-run speedup was byte-identical on 12 setups.

• A boundary that is easy to get wrong

  The equity loader is half-open (the session date is excluded) and the options path is inclusive. One function, engine_end(), is the only place that difference is expressed, so callers ask by mode instead of reasoning about it.

1. 01

   A covered call with the wrong sign

   Sixteen named strategies were run through the full engine on real prices and checked against the payoff identity a broker publishes for each. The first pass was 15 of 16. A weekly covered call reported a max profit of -51,286 where the published identity gives +9,964. Three places decided "does this structure span two expiries" by hand-testing one slot name; adding a third slot satisfied none of them. The fix is one predicate used everywhere, plus a test that fails if anyone writes the check by hand again. Code review had not caught it. Running the formula did.

2. 02

   The same strategy, different answers

   The default end date was evaluated once, at import time, so any run that did not name an end date was silently cut off at the API worker's boot date. Two workers restarted on different days gave different answers to the same request. The frontend also sent a UTC date, so between 00:00 and 05:30 IST it asked for yesterday. This is what the pin and fingerprint above exist to prevent.

3. 03

   A breakout that could never fire

   The Donchian channel had no one-bar shift, so it contained today's own high and "close above the upper band" was impossible: 0 fires in 12,275 bar-observations, which made three shipped breakout presets dead. After the fix it fires 198 times on RELIANCE alone. The same audit corrected 27 of the 62 indicators.

4. 04

   Indicators that started trading too early

   Several indicators emitted values from bar 0, so a "200-period" line was computed from 3 bars. Cold-start error reached 233% on ADX and 2,692% on TRIX. Because the engine loads no warm-up, bar 1 is a live trading decision: 15 of 218 trades on a measured ADX strategy entered before ADX was warm, six of them on the second bar. A +DI reading of 148.5 (a percentage above 100) came from numerator and denominator seeded from different bars. RSI returned NaN at exactly 100, so "RSI above 70" was false in the most overbought state possible.

5. 05

   A test suite that could not fail

   Mutation testing removed the %D smoothing from the stochastic oscillator entirely, and all 641 tests still passed. The indicator library had no arithmetic tests at all. It does now, and the practice stuck: new guards are mutation-checked, and several test drafts that turned out to be vacuous were thrown away.

6. 06

   A 2016 entry chosen with 2026 financials

   The first fundamental filter took each company's latest filing and applied it across the whole backtest window, which is the strongest form of look-ahead there is: "which companies screen as cheap and profitable today" is very close to "which companies did well", so testing whether they did well is nearly circular. The fix re-decides eligibility every quarter from filings published by that day. A column stamped during a backfill would have shown filings bunched in 2026 with absurd lags, which is how the four unusable 2017 rows were spotted.

7. 07

   Two devices, two answers

   Production runs two gunicorn workers, each with its own module-level cache and its own refresh clock, and nginx load-balances with no session affinity. Forty requests to one endpoint returned two distinct as_of timestamps, split 26 and 14, which is how a laptop and a phone showed different numbers for the same page. One shared snapshot in Redis fixed it: 2 distinct snapshots before, 1 after.

8. 08

   Days the exchange was shut

   The options equity curve was built on every business day. Measured on a 2024 to 2025 run, 30 of 523 points (5.7%) were exchange holidays, each a flat zero-return day that never happened, which understated volatility and moved Sharpe. The curve now uses real trading sessions.

9. 09

   Margin that was 11 to 15 times too light

   The first margin gate understated defined-risk structures against the broker's own SPAN calculator. Rebuilding it and measuring against that calculator brought the error to about 7%.

• Language in, validated rules out

  A forced tool call turns a sentence into the strategy schema, so the model returns a structured config validated by the same Pydantic schema as a hand-built strategy. A mode missing its inputs, or an impossible band, is refused with a plain sentence and never silently defaulted. The tool definition is sorted so it is byte-stable across restarts and the prompt cache is not invalidated.

• Prompt caching, accounted for

  Prompts use cache_control with 5-minute and 1-hour TTLs, and the spend ledger prices cache writes (1.25x and 2x) and cache reads (0.1x) separately. An earlier version left the cached route unrecorded, so spend was understated and the gate saw only a fraction of the bill; that was found and fixed.

• A spend gate, not just a ledger

  A daily ceiling for the whole platform and another per signed-in user turn the ledger into a gate. Sums are cached for a minute, so the worst case is a small overshoot, not a runaway bill.

• Bounded by design

  No recommendations, no price targets and no promised returns, in any phrasing. The product is analysis only: no broker connection and no order routing.

• v5

  What changed

  An end-of-window position is no longer charged a sell-side cost it never paid, so the equity trade log now sums exactly to the reported final value.

• v6

  What changed

  The options equity curve is built on real trading sessions instead of every business day (see holidays above). Sharpe moves; trades, P&L, CAGR and drawdown depth do not.

• v7

  What changed

  The indicator audit: 27 of 62 indicators corrected across warm-up, the Donchian shift, Wilder smoothing consistency and canonical bounds.

• v8

  What changed

  Instrument slippage is tiered by measured liquidity instead of one rate for every ETF, InvIT and REIT. Equity results proven byte-identical on four shapes.

• v9

  What changed

  An options leg whose exact strike did not trade now uses the nearest strike that did. Measured on 6 shapes and 845 trades from 2016 to 2026: no existing trade changed, and previously skipped cycles (3 of 126 on a monthly 5% OTM put-write, including 12 Mar 2020) are now filled.