Tradetron · FastBT — now the only backtest engine on Tradetron

We rebuilt our backtest engine from zero.
Here is the proof.

FastBT is a new simulation engine written for one contract: predict what a live deployment actually does — in seconds, on Indian, US and crypto markets, for any strategy the Strategy Builder can express. The highlights are on this page; every claim opens into its full evidence — measured runs, published audits, validation against real executed fills. Since 25 August 2026, every backtest on Tradetron — web or AI chat — runs on FastBT.

New pricing ₹20 per 6 months tested ₹20 flat per backtest — full 6+ year history, walk-forward included
~30 s
4-year multi-leg positional backtest, end-to-end incl. full report [1]
up to 486×
faster than the previous engine on identical jobs, measured [2]
6 segments
NSE F&O · cash equity · BSE · MCX · crypto · US markets (SPX & more)
274.9 M
option bars audited exchange-vs-store — zero OHLC violations [3]
60+
metrics per report, Monte-Carlo & walk-forward included
100%
of Tradetron backtests now run on FastBT — one engine, every user, since 25 Aug 2026 [9]
₹20
flat per backtest — full available history, any period, walk-forward free
First, the elephant

Our old backtest earned its criticism. We read all of it.

Backtests stuck in queues for days. History that started in 2020. No MCX, no crypto. Keywords that worked live but not in backtest. A trader on X summed it up:

“If I am using Tradetron, I should use the BhagwanBharose Platform.”

— @khanzubare, X (Twitter), June 2024. Fair, at the time.

FastBT wasn’t a patch on that engine. It is a replacement — built and instrumented so every one of those complaints now has a measurable answer.

That replacement is now complete: on 25 August 2026 the previous engine was retired and FastBT became the only backtest engine on Tradetron — for every user, on every plan. Before flipping, we replayed 306 real user jobs on the new fleet and checked every outcome — not one would have been lost. In that same sample, one in five previous-engine runs had quietly delivered nothing at all: no fills, no report. FastBT’s contract is fail-closed — every run either delivers, or tells you so and refunds. [9]

Shipped, not roadmapped

Every line below is live. Most of it came from a complaint.

A backtest engine earns trust the way a broker does — by being corrected in public and fixing it. These landed in the last fortnight. Each one was reported by someone, measured, fixed, and verified against raw exchange data before it shipped.

Custom Python runs natively Basket / list strategies trade every member The 5-year window cap is gone 86 more cash symbols & ETFs MCX copper · zinc · aluminium · silver ₹0 brokerage default Four missing market days rebuilt One-day reports deliver Option Greeks computed natively in-engine Refused runs refund automatically AI-chat backtests run on the same fleet
Speed

Hours became seconds. Measured, not marketed.

Same strategies, same windows, same data — old engine vs FastBT, timed on production hardware. The grey bar is the old engine; the indigo sliver is FastBT, on the same linear scale. The sliver is the point. And beyond hand-picked jobs: before the one-engine cutover we replayed 306 real user backtests on the production fleet — the median job ran 3.7× faster and the slowest tenth 42× faster, same template, same window, same fills [9].

Backtest windowOld engineFastBTSame job, side by side
18 days · 1-min35 min4.3 s
4.3 s
486× faster
6 months · 5-min3 h 21 m31 s
31 s
393× faster
2 months · 1-min3 h 35 m43 s
43 s
299× faster
3.7 years · hourly10 h 48 m5 m 29 s
5 m 29 s
118× faster

Paired runs of identical jobs on the production backtest fleet, August 2026 — cold-cache, contended-box conditions, i.e. understated [2].

Coverage

Every market. Every strategy. One engine.

Six segments on minute-level data — and if the Strategy Builder can express it, FastBT can test it: the same conditions, keywords, legs and exits your live deployment runs.

NSE index options BSE · SENSEX & BANKEX Stock options & cash equity MCX commodities Crypto · BTC & ETH US markets · SPX & more India VIX regime data
Short straddleIron condor 920 straddle w/ re-entryWeekly put ladder Supertrend directionalEMA-band spreads Gamma-scalped straddleCovered calls MCX crude condorBTC daily strangle SPX 0DTE strangle Cash-equity momentum basketsVIX-gated strangles Custom Python
The integrity layer

Scored against reality — not against another simulator.

Any backtester can agree with itself. FastBT’s data is audited exhaustively, its fills are scored against real executed trades from live deployments, and its reports are built to stop you fooling yourself. Each card opens the full evidence.

The report

Not a P&L number. A full diligence file.

Every run produces an interactive report — 60+ metrics across 20+ sections: verdict & grade, equity and drawdown, Monte-Carlo ranges, VIX-regime splits, leg and exit attribution, trades superimposed on the real chart, a live cost lab, and every raw fill. Each metric has a plain-English tooltip written for traders, not quants. Open a live sample report ↗

Two real samples, both generated by the production engine: a single 6.5-year run ↗  ·  a 12-variant sweep with walk-forward ↗ — the sweep one carries a variant navigator, so you can step through all twelve inside the same report. More on that below.

Sweep & walk-forward

Tune it — then find out whether the tuning was real.

Change one number and a strategy’s result changes. So test the number, not your hunch about it. A parameter sweep re-runs the whole backtest once per value and lays the results out as a grid. Walk-forward then splits the history, so the strategy is judged on data it was never tuned on. A sweep without walk-forward is just a machine for finding the luckiest number — which is why walk-forward is free here, and why the winner is chosen by it.

The sweep

Up to 4 parameters, up to 10 values each, up to 16 combinations in one run. Each cell is a full backtest — real charges, your slippage, the same fail-closed refusals — never an interpolation. Click any cell to open that variant’s own complete report.

The split

Choose an out-of-sample share (up to 90%; 30% is the usual pick) and the window splits in two: tune on the early part, verify on the rest. Every variant gets a plain verdict — holds up, degrades, fails out-of-sample — from how much of its in-sample daily average actually survived.

The winner

Not the best in-sample number — the best in-sample number that also holds up. If nothing holds up, the grid says so and labels the pick least-bad, not deployable instead of crowning it. Variants that looked strong in-sample and collapsed after are flagged ⚠, not quietly ranked.

Here is a real one, and it is not flattering. Twelve entry/exit-time combinations of a NIFTY intraday short straddle over 18 months, 30% of the window held back. In-sample they look like a menu of winners — +6.5% to +21.8%, every one profitable. Out of sample, ten of the twelve lose money and the other two merely degrade. The best in-sample variant, +21.8%, comes back at −3.3% on data it was not tuned on. Nothing holds up, so the report crowns nothing and says in writing: treat the whole grid as overfit. Twelve ways to be wrong, found for ₹240 instead of in a live account. Open the sample sweep report ↗

A sweep costs 1 credit per variant — a 6-variant sweep is ₹120, and the number is on screen before you run it. Walk-forward adds nothing, on a sweep or on a single backtest.

AI-native · nobody else has this

Backtest without opening Tradetron at all.

Tradetron ships an MCP server — the open standard that lets AI assistants use real tools. Connect Claude (or any MCP-capable AI) to your account once, describe a strategy in plain English, and get a real editable template and the full backtest report URL back in the chat, seconds later. No other algo platform — Indian or global — offers this.

See it working: One idea, thirteen versions ↗  ·  How to connect your AI ↗

The comparison

Where FastBT stands. Check every cell.

Compiled August 2026 from each platform’s public documentation, pricing pages and user forums. If we got a cell wrong, tell us and we’ll fix it.

vs Indian retail backtesting platforms

CapabilityFastBT
(Tradetron)
Platform APlatform B Platform CPlatform DPlatform EPlatform F
NSE index options, minute-level 2020→, audited bar-by-bar 7.5 yrs, 1-min ~ weakno true option-leg BT 7 yrs on top plan ✗ EOD only
Stock options & cash equity F&O stocks + 1,700+ cash symbols ~~50 stock options, no cash ✓ equity ~ ~ 51 stocks ~ 7 stocks
MCX commodities backtest futures + options, evening session ~ indicators only their own FAQ
Crypto backtest (BTC/ETH) options + perps, 24/7
US options backtest (SPX/SPY/QQQ) 0DTE + weeklies, validated vs live US fills
Multi-year positional backtests 4 yrs in ~30 s ~~3 months on 5-min candles ~ EOD metered by the minute
Custom Python in backtest same code as live deploy ✗ no-code only ₹5,310+/mo tiers
Parameter sweep / optimiser 16 variants, up to 4 axes portfolio combiner only ~ Greeks tests
Walk-forward / out-of-sample ✓ freeverdict on every run
Monte-Carlo robustness in report
Margin modelling in report peak/avg margin, return on margin ~ Greeks/payoff
Historical lot sizes per trade date from exchange circulars ?? ??
Validated against real live fills 21,011-fill ground-truth program
Create + backtest from inside an AI (MCP) report URL delivered in the chat
Backtest → one-click paper deploy same template, same engine keywords via bridge software
Backtest pricing ₹20 flatany period · WF included ₹1/run · ₹1.5k/mo25 free/week freebroker clients only ₹1.3k–5.4k/mo ₹1k/mo + demat ₹1.1k/30d ₹1.2k–23.6k/mo
supported ~ partial / restricted ? not documented not available

Sources: each platform’s public documentation, pricing pages, and public user forums as of August 2026. Platform names withheld — the cells stand on their own. Cells marked “?” were not documented publicly.

Why this matters

Deploying an untested idea isn’t trading. It’s donating.

A backtest is the only way to confront an idea with thousands of days it didn’t get to choose. And the engine you test on matters as much as testing at all: repainting, fabricated ticks or silent data holes don’t reduce your risk — they manufacture false confidence. That is why FastBT’s integrity layer exists.

Pricing

One flat price. No meters, no traps.

₹20per backtest — that’s it.

Compare: unlimited plans elsewhere run ₹1,499–₹5,417/month whether you test or not; code-first platforms meter by the minute; one platform requires opening a demat account just to backtest. ₹20 to interrogate an idea against six years of audited data is the cheapest insurance in trading.

Next steps

Three ways in. Pick one.

1 · Read a real report

Open the 6.5-year NIFTY weekly put-ladder sample ↗ — the full diligence file including the new trades-on-chart view — with the section-by-section walkthrough above as your guide.

2 · Watch one AI session build a strategy

One idea, thirteen versions ↗ — a real, unedited journey of a put-ladder strategy going from a one-line idea to a backtested, deployable template in a single Claude session, every version backtested along the way.

3 · Connect your own AI

Connect Claude or any MCP-capable assistant ↗ to your Tradetron account — then describe a strategy in plain English and get the backtest report URL back in the chat.

FastBT

Stop trusting. Start verifying.

Describe a strategy in the builder — or in plain English to your AI — and have a full diligence report in your hands before your chai gets cold.

Run a backtest See a sample report

What shipped — and how we checked it

Nothing here is a roadmap item. Each was released to the production engine, regression-gated against byte-identical control runs, and — where it touches prices — checked fill-by-fill against the raw exchange store.

Custom Python, on the fast engine

Strategies with your own Python used to fall back to the old engine. The cause was a routing rule gated on a setting that was never switched on anywhere, so it could never be satisfied. The engine had been able to run Python for weeks. Removed — and the same code your live deployment runs is now what the backtest runs.

Fewer strategies turned away

Six families of building blocks that used to be refused now run — second-level timing, random/month/year values, state-carrying keywords, instrument-based exits and multi-symbol lists. Measured across 323 live-traded strategies and 111,009 real fills: the share of real trading volume we cannot honestly backtest fell from 44% to 34%, with zero regressions.

Baskets trade every member

A strategy on a saved list used to collapse to a single stock. Now a 50-member NIFTY 50 basket books 1,118 fills in 12.9 s, and entry-side parity against a real ledger came back 7 of 7 exact to the paisa. Delisted and renamed members no longer kill the run. Exits on baskets are still being validated — we say so rather than imply otherwise.

ETFs and 86 more cash symbols

GOLDBEES, SILVERBEES, PHARMABEES, ITBEES, MON100, CPSEETF and 80 more used to return “unknown underlying”. Every symbol was checked for real minute bars over 400 days before it was registered — one was rejected for having none. A silver/gold ETF rotation now books 62 fills, each price matched to the exchange bar.

MCX base metals and silver

Copper, zinc, aluminium and Silver-100 now backtest. A three-metal futures strategy books 284 fills spread across all three — so no leg silently no-ops — on real front-month contracts that roll correctly. ZINC 31JUL S 5000 @ 364.10 matches the store bar close exactly.

Your whole history in one run

A submit-time cap quietly truncated long requests to five years. It is gone — the engine itself never had one. Test a weekly straddle across the March-2020 crash and the 2022 drawdown in a single job instead of stitching two.

Costs you recognise

Brokerage now defaults to ₹0 and slippage to 0.05% (was 1%). We also found the engine and the report were running two different cost models — on one 25,948-fill job they disagreed by ₹637,241. Aligned to 0.19%. The risk metrics were the worse casualty: one job’s Sharpe fell 4.43 → 0.48 once costs were actually in the series. Better a true 0.48 than a flattering 4.43.

Four missing market days

A user reported his exits not firing on a Tuesday monthly expiry. Four entire NSE trading days were missing from our store. Rebuilt from the official exchange bhavcopy — 257,959 points, all seven data nodes. The hole had force-settled 654 legs the next day, many at garbage prices, which exposed a second bug: a worthless expiring leg was being booked at its entry price instead of zero. Both fixed. [3]

One-day reports arrive

A one-day backtest computed its result and then died building the report — silently, with no refund and no email. The signature was absolute: no single-day backtest had ever been delivered. Root-caused to one line, fixed, and proven by delivering one end-to-end.

Three worked examples

Real runs on the production engine, with the fill counts they actually produced. Every price below was checked against the raw exchange bar for that minute — not against another simulator.

A 50-stock NIFTY 50 basket

1,118 fills · 12.9 s. A swing strategy over every member of the index. Entry-side parity against a real trade ledger: 7 of 7 exact — same date, same member, same quantity, same price to the paisa. Before this, the same strategy traded one stock and you would never have known.

Copper + zinc + aluminium together

284 fills — COPPER 90, ALUMINIUM 98, ZINC 96 — so you can see no leg silently did nothing. Contracts roll on the real calendar (31 Jul → 30 Sep), never a phantom expiry. Spot check: ZINC 31JUL S 5000 @ 364.10 equals the exchange bar close for 10:00 IST that day.

A silver / gold ETF rotation

62 fills. A SILVERBEES–GOLDBEES rotation — the kind of strategy that returned “unknown underlying” a fortnight ago. Spot check: SILVERBEES B 250 @ 231.79 at 10:15 IST equals the bar close of 231.79. Both symbols were verified to carry 400+ days of real minute bars before we enabled them.

What you said → what FastBT does now

What you said
What FastBT does now
“My backtest was in the queue for three days.”
Seconds, not days. A 4-year positional options backtest completes in ~30 s end-to-end, report included. A 1-year intraday straddle: ~14 s. [1]
“Backtest data only goes back to 2020, and there’s no MCX or crypto.”
6+ years of minute data on NSE, plus BSE (SENSEX/BANKEX), MCX commodities, BTC/ETH crypto, and now US markets — segments the old engine could not backtest at all.
“It picked the wrong strikes. The result had no relationship to the strategy.”
Strikes resolve against the actual traded chain for that date, with per-date lot sizes from exchange circulars — and the engine is scored against real executed fills from live deployments. [4]
“It returned 15 days of a 2-year request and still burned my credits.”
Reports are fail-closed. Every report must reconcile against its own trade ledger before it is allowed to publish. A zero-trade run says so explicitly instead of pretending.
“Paper showed profit, live showed loss. Same strategy, same day.”
We measure that gap instead of ignoring it. FastBT is continuously validated against 21,011 real fills from live deployments — the only Indian platform that scores its backtester against reality, not another simulator. [4]
“Run the same backtest twice, get two answers.”
Deterministic by construction. The same job run on four different machines produced byte-identical fill books, five out of five runs. [5]
“My exits didn’t fire on the Tuesday monthly expiry.”
He was right, and it was worse than his one strategy. Four entire NSE trading days were missing from our historical store. We rebuilt all four from the official exchange bhavcopy across all seven data nodes — 257,959 data points — and fixed the settlement bug the holes had exposed, where a worthless expiring leg was booked at its entry price instead of zero. [3]
“My strategy uses custom Python, so it always fell back to the slow engine.”
Custom-Python strategies now run on FastBT. The block was a routing rule that depended on a setting which was never switched on anywhere — so the condition could never be true and every such job quietly went to the old engine. Removed; the same Python your live deployment runs is now what the backtest runs.
“My basket strategy only ever traded one stock.”
Saved symbol lists now trade every member. A 50-member NIFTY 50 basket books 1,118 fills in 12.9 s. Entry-side parity against a real ledger came back 7 of 7 exact — date, member, quantity and price to the paisa. A delisted member used to kill the whole run; renames are now stitched.
“It said no trades. My strategy definitely trades.”
Sometimes it was our engine failing, not your strategy. That verdict was decided from a single signal — an empty position table — so an engine failure and a genuinely quiet strategy looked identical. We re-ran a labelled sample to measure how often it was us, and the fix that separates the two provable cases is built and in review.
“A one-day backtest never arrived.”
Fixed. A one-day run produced its result and then failed while building the report, silently. The separation was absolute — no single-day backtest had ever been delivered. Root-caused, fixed, and proven end-to-end by delivering one.
“The costs in my report don’t look like my broker’s.”
Defaults are now honest, and you control them. Brokerage defaults to ₹0 — most users are on zero-brokerage brokers — and slippage to 0.05% instead of 1%. We also found the engine and the report were using two different cost models and aligned them; on one 25,948-fill job they had disagreed by ₹637,241.
“A stop at 90 should trigger when the candle’s low touches 90 — not only when it closes below.”
Correct, and it is built. Stops and targets can now be evaluated against the bar’s true high and low rather than a single price. It is not switched on yet, deliberately: when one bar’s low hits your stop and its high hits your target, minute data cannot say which came first, and the safe assumption makes results look worse. We would rather ship that with the choice explained than flip it silently. built · not yet live

How it’s this fast

The engine parallelises across days and independent position cycles, bulk-loads minute data, and evaluates conditions vectorised instead of bar-by-bar. More hardware makes it faster automatically; no code change required. All timings from paired runs of identical jobs on the production backtest fleet, August 2026 [2] — cold-cache, contended-box conditions, i.e. understated.

Long positional? Also seconds.

The classic killer — a 4-year positional options strategy — runs in ~30 s end-to-end including the full analytics report, and ~14 s with a warm data cache. The old path took 2–6 hours on comparable windows, when it finished at all. [1]

Parameter sweeps at the same wall-clock.

A 4-variant, 12-month sweep finishes in the same ~14 s as a single backtest — variants run concurrently under a core budget. A 16-variant grid completes in about two minutes. [6]

Speed that survives an audit.

Every optimisation shipped only after producing byte-identical fill books against the slow path. When one cache shortcut was found stealing correctness, it was reverted and the engine took the 38% slowdown. Speed is never allowed to buy wrong answers.

Market-by-market detail

NSE index options

NIFTY, BANKNIFTY, FINNIFTY — weekly and monthly expiries, full option chains, minute bars back to 2020, with the real expiry-day migrations (Thu→Wed→Tue) handled per-date.

BSE index options

SENSEX and BANKEX chains — with per-circular lot histories, including the 2026 BANKEX spot rename stitched correctly. Most Indian platforms added BSE late or not at all.

Stock options & cash equity

F&O stock options plus 1,300+ cash-equity symbols — CNC equity strategies, list-based baskets, and screener-driven entries across the stock universe.

MCX commodities others: absent

Crude, Gold, Silver, Natural Gas, Copper and more — futures and options, with evening sessions (09:00–23:30), commodity expiry calendars, and CTT in the charge stack. No listed Indian retail competitor backtests MCX options. [7]

Crypto others: absent

BTC & ETH on Delta India — perpetuals and daily/weekly/monthly options, true 24/7 sessions, USD accounting, fractional coin lots. Backtest the strategy you actually deploy on crypto.

US markets new

SPX, SPY, QQQ, NDX, XSP option chains — 0DTE and weekly expiries on minute-level data, plus US index spots. Scored against real fills from live US deployments: entry minutes and strikes matched exactly.

Regime data

India VIX as a condition input — and every report breaks performance down by VIX regime, so you see how the strategy behaves in calm vs stressed markets before you deploy into the wrong one.

One engine, one report format — whether the strategy trades a NIFTY straddle, a crude oil condor, a cash-equity basket, a BTC perpetual, or an SPX 0DTE strangle.

If the Strategy Builder can express it, FastBT can test it

FastBT isn’t a separate mini-language with ten templates. It executes the same Advanced Strategy Builder logic your live deployment runs — conditions, keywords, legs, exits and all.

Multi-leg options structures

Straddles, strangles, iron condors, iron flies, ratio & calendar-style spreads, hedged wings — with ATM re-selection per entry, strike offsets, premium-based and Greeks-aware selection.

Intraday and true positional

EOD-flat intraday, overnight, weekly and monthly carry, expiry rollovers, re-entry counters, universal exits — multi-year positional runs that other engines cap at months or refuse outright.

Indicator & price-action logic

100+ keywords: SMA/EMA crossovers, RSI, Supertrend, MACD, Bollinger, CPR, fractals, Renko and Heikin-Ashi candles, VWAP, OI & PCR, multi-timeframe conditions on any instrument.

Risk logic that actually simulates

Per-leg SL/TP, trailing stop-loss state machines, lock-and-trail profit, time-based exits, repair/adjustment legs (once and continuous), overall MTM stops.

Custom Python

Strategies with embedded python_code blocks run in backtest too — the full-flex tier that no-code-only backtesters simply can’t represent.

Baskets & lists

List-based strategies over multi-symbol universes — 55,000+ user lists resolvable — so basket and screener strategies backtest the same way they deploy.

A backtest is only as honest as its data. We audited ours — all of it.

Every platform claims “clean data.” We published the audit instead.

274,890,370 option bars, zero violations

Every strike, CE and PE, of all 599 NIFTY & BANKNIFTY expiries 2020–2025 was OHLC-validated bar-by-bar: 0 high<low, 0 out-of-range, 0 zero/negative prices, 0 duplicate timestamps. Exhaustive — not sampled. [3]

Holes found, then filled

The audit surfaced missing trading days and 141 per-strike ingest holes near ATM. They were repaired from source vendor data, re-verified across all seven storage nodes, and the closure sweep now reports zero flags across 413 expiries.

Ghost expiries deleted

When NSE moved expiry days, stale ingest rules minted phantom expiry series that could silently break chain resolution. All 22 ghost tokens were identified and removed — a defect class competitors haven’t even audited for.

Standing guard, not a one-off

A daily gap alarm asserts every trading day has data; a weekly automated regression sweep re-runs the strike-hole and ghost-expiry audits. Data quality is monitored, not assumed.

Real lot sizes, per trade date

NIFTY’s lot went 75→50→25→75→65 across the years. FastBT applies the actual lot size in force on each trade date from exchange circulars. Engines that apply today’s lot to 2021 misstate every position size by up to 3×.

Session-faithful

MCX evening sessions, crypto’s 24/7 clock, special Saturday sessions, half days and the 2021 NSE outage are all in the calendar — the engine trades the sessions that actually happened.

Scored against reality — the ground-truth program

Any backtester can agree with itself. The only test that matters is: does the engine predict what a live deployment actually did? FastBT is the only Indian backtesting engine with a standing program that measures exactly that.

Ground-truth validation

A corpus of 21,011 real executed fills from live auto-trading deployments — real money, real brokers, real slippage — is replayed through FastBT and scored on decision recall, precision, timing and fill-price error. Gaps it finds become engine fixes, not footnotes. [4]

Nightly fills audit

Every night, served backtest fills are sampled and re-verified against the raw market data store — is every fill priced within its own bar, slippage accounted, at a price that existed? Invariant violations page the team.

Fail-closed reports

Before any report publishes, an independent reconciliation gate re-derives the P&L from the trade ledger. If the stats don’t match the trades, the report is blocked — no email, no chart, no quietly-wrong number. This gate has caught legacy reports overstating trade counts 55–61×.

Deterministic & reproducible

Identical job → identical fill book, byte for byte, across different machines and repeated runs [5]. Engine builds are content-fingerprinted, so any result can be reproduced exactly.

Cross-checked leg-by-leg

On identical windows, an independent reviewer matched 4,514 legs (99.4%) to the exact minute and paisa against the previous engine — with every divergence root-caused and resolved in favour of the historically-correct convention. [8]

Loud about what it can’t do

If a strategy uses a construct FastBT can’t faithfully simulate, it says so and declines — it never silently books a wrong answer. Skipped signals and data-gap sessions are disclosed in the report footer.

Built to stop you fooling yourself

The biggest risk in backtesting isn’t bad data — it’s a good-looking curve that’s pure luck. The academic literature is blunt about this: try enough parameter combinations and a “winning” backtest is guaranteed, live performance is not.

Bailey, Borwein, López de Prado & Zhu (2014), “Pseudo-Mathematics and Financial Charlatanism,” Notices of the AMS: high simulated performance is easily achieved by trying a modest number of strategy configurations — and without knowing how many were tried, the result is unassessable.
Bailey & López de Prado (2014), “The Deflated Sharpe Ratio,” Journal of Portfolio Management: a reported Sharpe ratio must be discounted for selection bias, trials, skew and fat tails before it means anything.

FastBT builds the defences into the product:

Walk-forward, out-of-sample — free

Split the history: tune on the early part, verify on data the strategy never saw. The report renders a plain verdict — holds up / degrades / worse — on every run, at no extra credit.

Sweeps that pick honest winners

Parameter sweeps (up to 4 axes, 16 variants) select the winner as best in-sample that holds up out-of-sample — never the raw in-sample max, which is overfitting by construction. Fragile winners are flagged, not crowned. See the sweep grid ↓

Probabilistic Sharpe & Monte Carlo

Every report carries a Probabilistic Sharpe Ratio (is the Sharpe real, given trades and tails?) and a block-bootstrap Monte Carlo: terminal P&L at p5/median/p95 and a 95th-percentile drawdown — the honest range, not one lucky path.

How to read the report, section by section

Open the live sample report ↗ and read it in this order:

Verdict & grade

Start at the top: an A–F grade with the main reason named, plus the trust checks — Prob. Sharpe (could this be luck?), Walk-forward (did it survive unseen data?), Concentration (is the profit three lucky months?), and Worst rolling 12M (what would the unluckiest subscriber have lived through?).

Headline tiles

Net P&L (after brokerage, slippage and statutory charges — judge on this, not gross), ROI and CAGR against capital actually required (worst-case margin, not a made-up number), Sharpe, Sortino, Calmar, max drawdown, win rate, expectancy per trade.

Equity curve & drawdown

Gross vs net vs benchmark, drag-to-zoom. The gap between the gross and net lines is what costs eat. The drawdown pane under it shows depth and duration — how long you’d sit underwater, which matters more to real people than the max-DD number.

Monte-Carlo robustness

The daily return path is block-bootstrapped thousands of times. Read Terminal p5 (the bad-luck outcome), median, p95, and Max-DD p95. If p5 is a result you couldn’t stomach, don’t deploy — whatever the headline says.

Consistency: monthly & calendar views

Monthly returns as % of capital, a daily-returns calendar, and the daily distribution with skew, kurtosis, tail ratio, VaR and CVaR — is this steady income or a lottery ticket with a good year?

Regime & timing analysis

Performance by India-VIX regime answers “what happens when volatility spikes?” Day-of-week timing edge comes with t-stats, so a Tuesday “edge” without statistical significance is labelled as noise, not sold as insight.

Attribution: legs, exits, drawdowns

P&L by leg shows which side of the structure earns and which bleeds. Exit attribution reveals whether profits come from your planned exits or from stop-losses and expiry forcing your hand. Top-5 drawdowns name the exact episodes.

Trades on the chart New

Price & trade activity superimposes your trades on the actual underlying chart: daily buy/sell bars and trade-count bubbles ride the real spot line, with a monthly consolidated view. Click any day and the report fetches that day’s option legs’ own minute-by-minute premium paths, with ▲buy/▼sell markers on the exact fills — you see every trade in the market context it happened in, not as a row in a table.

Margin & capital reality

Peak and average margin the book actually blocks, and return on peak margin — the number that decides whether the strategy beats simply not trading. Plus Ulcer index, pain index, recovery factor, and time-to-recover.

Cost lab

Interactive sliders re-compute net P&L live as you change slippage and brokerage assumptions. If a strategy dies at one extra rupee of slippage per lot, you learn it here — not in month two of deployment.

Trades & raw fills

Full round-trip trade analysis (profit factor, avg win/loss, hold times, streaks) down to every raw fill with instrument, minute, price, lots and charges. Nothing is hidden — audit any number back to its trades.

Custom period re-slicing, week/day/underlying breakdowns, risk flags, and auto-generated “hypotheses to test” round out the file. Every metric has a plain-English tooltip written for traders, not quants.

The sweep grid, and what it refuses to claim

Every number below is how the shipped engine actually behaves — the limits, the verdict thresholds and the winner rule are the ones running in production. Open the sample sweep report ↗ alongside this.

You sweep what your strategy already contains

The parameters offered are the numbers found inside your own strategy — entry and exit times, ATM strike offsets, expiry offsets, indicator periods, lots — each named after the building block it sits in. On submit the server re-checks every value against that same list, so a value that is not in your strategy can never be swept. Two honest limits: 0, 1 and −1 are skipped (they are almost always flags, not tunables), and only literal numbers are visible — anything computed at runtime or held in a variable cannot be discovered this way.

The limits, and why they exist

4 axes × 10 values × 16 combinations maximum. The cap is not arbitrary: one combination is one complete backtest, so an uncapped 4×10 submit would be 10,000 jobs from a single click. The grid you get back is therefore small enough to read and expensive enough to think about — which is the point.

Every variant is a real backtest, in one real report

No interpolation, no shared state, no shortcut. Each variant runs the full engine on the full window with real brokerage, statutory charges and your chosen slippage. What you get back is a complete Tradetron backtest report — the same 60+ metrics, cost lab and trades-on-chart as a single run — with a variant navigator on top: step through all twelve, or read them at once as a sortable table or a heat grid, with in-sample and out-of-sample shown side by side and the walk-forward verdict on each row.

How the verdict is computed

The window splits at your out-of-sample percentage. Each side’s daily average P&L is measured, and the verdict is the ratio between them: holds up when out-of-sample keeps at least 40% of the in-sample daily average, degrades when it stays positive but below that, and fails out-of-sample when the sign flips. A variant that never made money in-sample is labelled IS negative rather than being scored at all.

How the winner is picked

Among variants that hold up and were profitable in-sample, the highest in-sample net wins. Picking the raw in-sample maximum is overfitting by construction, so it is not what happens. If nothing holds up, there is no crown — the grid returns the least-bad variant and says in writing that it is not a deployable pick. The ⚠ flag marks any variant that looked good in-sample (ROI above 3%) and then degraded, flipped or was negative to begin with.

What it refuses to claim

If a variant comes back with an empty ledger, the grid does not call it “no trades”. It is shown as not compared, excluded from the ranking, and the winner is downgraded to provisional with the reason printed on the report. A zero-fill result is not proof your strategy makes no trades at that value — it can equally be the engine missing them, and we have had that failure. The report says only what it knows and tells you to re-run that value on its own to tell the two apart.

What it costs

1 credit — ₹20 — per variant, because each variant is a full backtest. A 6-variant sweep is ₹120; the 16-variant maximum is ₹320. The total is quoted in the request panel before you submit. Walk-forward is free: the out-of-sample split is never a price multiplier, on a sweep or on a single run. And the window is free too — one day or the full history, the price per backtest is the same.

The sample grid is a genuine production run, published unedited — twelve combinations that all looked profitable in-sample and all lost out-of-sample. We publish that one on purpose: a sweep tool whose samples always find a winner is selling you the overfitting it claims to prevent.

The idea-to-evidence loop, inside your AI

Describe the strategy in plain English

“Short a NIFTY strangle at 9:20, 25-delta strikes, 25% SL per leg, exit 15:10, re-enter once.” No builder, no code, no forms.

The AI builds it in your account

The assistant creates a real, editable template in your Strategy Builder — validated against the same keyword catalog your live deployment will run. You can open it in the ASB and inspect every condition.

Ask for the backtest — get the report URL in the chat

The AI runs FastBT and hands you the full interactive report link right in the conversation, seconds later. Ask follow-ups — “tighten the SL and re-run”, “sweep the entry time” — and iterate at the speed of conversation.

Deploy to paper when it earns it

Same template, same keywords, same engine semantics — one click from backtest to paper trading. The idea-to-evidence loop closes without leaving your AI.

No other algo platform — Indian or global — offers an MCP integration at all, let alone create-and-backtest from inside the AI. See it working in a real session: One idea, thirteen versions — building a strategy with Claude ↗  ·  How to connect your AI to Tradetron ↗

vs global platforms — why they can’t serve Indian F&O

CapabilityFastBTGlobal platform AGlobal platform B Global platform CGlobal platform DGlobal platform E
NSE/BSE options data, ready to use ✓ built in no option chains India equities only ✗ BYO feed~365 d of 1-min via vendors no NSE brokers CME/ICE/Eurex only
Indian multi-leg options backtest strikes, expiries, lots native ✗ impossiblesingle-symbol tester US options only ~ DIYhand-rolled strike logic no options class
MCX commodities ~ charts only ~ BYO
No code required builder UI or plain English via AI ~ proprietary script ✗ Python/C# ✗ proprietary language ✗ proprietary language ~ + C#
Documented realism failures none in the fill pathknown gaps fail closed, with a named reason repaintingdefaults 30–50% optimistic slow queue strike-freeze trap fabricated ticks OHLC fill guessing
Cost to backtest Indian options ₹20 / run $59.95/mo…and still can’t $10–96/mo…and still can’t $279 + ₹500/mo feed free…and can’t at all free…and can’t at all
supported ~ partial / restricted ? not documented not available

Sources: each platform’s public documentation, pricing pages, and public user forums as of August 2026. Platform names withheld — the cells stand on their own.

Six questions to answer before real money touches a strategy

Every strategy idea feels right in your head — that’s why you had it. A backtest is the only way to confront the idea with thousands of days it didn’t get to choose. Before real money touches a strategy, you should be able to answer:

Does the edge exist at all?

Not “was last month good” — did it work across 6 years, both market regimes, hundreds of expiries? Most ideas die here, cheaply, at ₹20 instead of ₹2 lakh.

Does it survive costs?

Brokerage, statutory charges and slippage kill more retail strategies than bad signals do. FastBT charges every simulated trade the full stack and lets you stress it further in the cost lab.

Can you live through its worst stretch?

The max drawdown, the longest underwater spell, the worst rolling year — if you’d have quit at the bottom, the strategy’s average return is irrelevant. The report shows you the bottom before you buy it.

Is it skill or luck?

Probabilistic Sharpe, Monte-Carlo ranges and walk-forward verdicts separate a real edge from a fitted curve — the exact discipline the academic literature demands and almost no retail platform ships.

How much capital does it truly need?

Peak margin, not a guess. Under-capitalised deployment turns survivable drawdowns into forced exits at the worst price.

What breaks it?

VIX-regime splits, leg attribution and exit attribution tell you the conditions under which the strategy stops working — so you can stop it first.

The receipts — where every number comes from

  1. 4-year (2021–2024) multi-leg positional options strategy, end-to-end through the live service including report generation: 27–52 s cold cache, ~14 s warm. 1-year intraday NIFTY straddle: ~14 s. Measured July–August 2026, production hardware.
  2. Paired runs of identical jobs, previous engine vs FastBT, on the production backtest fleet (Aug 2026): 2,108 s → 4.3 s; 12,060 s → 30.7 s; 12,917 s → 43.2 s; 38,857 s → 328.9 s. Cold-vs-cold on a contended box — conservative.
  3. Exhaustive OHLC audit of the backtest data store: 274,890,370 option bars across all 599 NIFTY + BANKNIFTY expiries 2020–2025, every strike, CE and PE — zero violations. Full audit report published; repairs and weekly regression sweeps documented.
  4. Ground-truth corpus: 21,011 executed fills from 12 live auto-trading deployments (39–63 trading days each), replayed and scored continuously; wider corpus of 111,037 real fills across 324 live-traded templates used for coverage scoring.
  5. Determinism: identical crypto options job submitted 5× across 4 different fleet machines — identical books each time (2,568 fills, 8 sessions), August 2026.
  6. 4-variant × 12-month sweep: 37 s serially, 14 s parallelised — equal to a single run’s wall-clock. Scales with cores.
  7. Competitor MCX/crypto absence per their own public documentation and FAQs, August 2026.
  8. Independent leg-by-leg reconciliation on identical 4-year windows: 4,514 of 4,543 legs matched on instrument, minute, direction and price (99.4%); zero price discrepancies among matched legs; residual differences root-caused to lot-size conventions, resolved to per-date historical lots.
  9. One-engine cutover (25 Aug 2026): before retiring the previous engine, 306 real user jobs from the prior five days were replayed byte-for-byte on the production FastBT fleet — none would have been lost to the switch. On validated matched pairs (previous engine demonstrably delivered fills), the median job ran 3.7× faster and the 90th percentile 42× faster. One in five previous-engine runs in the sample had recorded a completed-looking job with zero fills and no report delivered.