Benchmarks
Backtesting speed benchmarks
Most backtesting platforms publish no measured speed. On 2026-10-31 we looked at 18 other platforms and engines; 8 gave a figure of some kind, and only a few state the hardware, data and test size behind it. AlphaPrime measured about 1.2 billion strategy evaluations in about 18 hours on one RTX 3090, using 10-Year Treasury Note futures and 37-minute bars. These numbers count different work, so they cannot rank the platforms: a bar processed is not a strategy evaluated. Each row names its conditions and links to its source. Where we found nothing, the row says so.
20 platforms and reports, all checked 2026-10-31.
Platforms with a published figure
The figures are copied as each source states them. Nobody here re-ran them, apart from our own row.
| Platform | Published figure | Conditions stated | Type | Source |
|---|---|---|---|---|
| AlphaPrime | About 1.2 billion strategy evaluations in about 18 hours. | One NVIDIA RTX 3090 (24 GB). 10-Year Treasury Note futures, 37-minute bars, 2015-01-01 to 2026-09-25, 114,371 bars per evaluation, fees and slippage included. Test date 2026-10-05. | Our own test | AlphaPrime benchmark page Checked 2026-10-31 |
| AmiBroker | An exhaustive optimization of 100 steps takes 25 seconds. The AFL language can process up to 166 million data bars per second. | Optimization: 10 years of end-of-day data, 100 symbols, hardware not stated. AFL figure: 2 GHz CPU. | Vendor figure | AmiBroker features page Checked 2026-10-31 |
| VectorBT PRO | Sparse-order simulation at 4.7 billion bars per second (low level) and 320 million (high level). Backtrader ran the same workload at about 0.12 million bars per second. | Apple M3, 8 cores. 1-minute bars, 1,000,000 bars in each of 8 columns, 1,388 filled orders per column. The vendor publishes the test suite. | Vendor figure | VectorBT PRO benchmarks Checked 2026-10-31 |
| RaptorBT | About 13 million bars per second: 25 million bars in 1.92 seconds on one SMA crossover backtest. | Apple M4, 10 cores, 24 GB. Random-walk prices, fastest of several hundred repetitions. | Vendor figure | RaptorBT wiki Checked 2026-10-31 |
| backtrader | 14,713 candles per second on CPython and 34,971 on PyPy. | Windows 10 laptop, i7, 32 GB. 100 files of 20,000 candles each, 2 million in all. | Vendor figure | Author's blog, 2019 Checked 2026-10-31 |
| QuantConnect LEAN | A live chart of data points per second for each benchmark algorithm. Its axis ran up to 2 million on the check date. | Master branch, more than 20 benchmark algorithms. Hardware not stated on the page. | Vendor figure | LEAN performance benchmark Checked 2026-10-31 |
| StrategyQuant X | Thousands of backtests per second and tens of thousands of strategies per hour. | Depends on data and test precision, per the vendor. Hardware, instrument and timeframe not stated. | Vendor figure | StrategyQuant features page Checked 2026-10-31 |
| StrategyQuant X | A user reported 457,874 strategies per hour; a forum administrator reported about 150,000 per hour. | User: 28-core i7, 32 GB, default builder settings. Administrator: 8-core Xeon E-2286M. Instrument and timeframe not stated. | User report | StrategyQuant forum thread Checked 2026-10-31 |
| MetaTrader 5 | With the MQL5 Cloud Network, calculations that would take a few days can finish in one hour. | Workload, pass count and hardware not stated. | Vendor figure | MetaTrader 5 help Checked 2026-10-31 |
| Adaptrade Builder | A testimonial reports a population of up to 65,000 on daily data finishing in 15 minutes. | i7-980x, 12 GB, SSD. Instrument not stated, undated. | User report | Testimonial on the vendor page Checked 2026-10-31 |
A vendor figure is what the vendor says about its own product. Do not divide one row by another: they measure different things on different machines.
Platforms with no measured figure
We read the pages linked in each row. If a figure sits on a page we did not reach, write to us and we will add it.
| Platform | What we found | Conditions stated | Type | Source |
|---|---|---|---|---|
| Build Alpha | No measured figure. The home page says thousands of strategies in under one minute, without hardware, data or test size. | Not stated | No figure | Build Alpha home page Checked 2026-10-31 |
| MultiCharts | Not found. The home page has no figure, and the product and forum pages blocked automated reading on the check date, so this needs a second look. | Not stated | No figure | MultiCharts home page Checked 2026-10-31 |
| TradeStation | Not found on the desktop platform page. | Not stated | No figure | TradeStation Desktop page Checked 2026-10-31 |
| NinjaTrader | Not found on the home page. The backtesting pages we tried would not open. | Not stated | No figure | NinjaTrader home page Checked 2026-10-31 |
| MetaTrader 4 | Not published in the strategy tester help. | Not stated | No figure | MetaTrader 4 tester help Checked 2026-10-31 |
| TradingView | Not published. The Strategy Tester section uses words only. | Not stated | No figure | TradingView features page Checked 2026-10-31 |
| vectorbt (open source) | No measured figure. The site says many thousands of strategies in seconds, without a test size or hardware. | Not stated | No figure | vectorbt home page Checked 2026-10-31 |
| backtesting.py | No measured figure. The site says hundreds of variants in seconds, without a test size or hardware. | Not stated | No figure | backtesting.py home page Checked 2026-10-31 |
| Zipline Reloaded | Not published on the repository page. | Not stated | No figure | zipline-reloaded repository Checked 2026-10-31 |
| NautilusTrader | Not published in the README or the docs we read. They describe the Rust core as fast, with no measurement. | Not stated | No figure | NautilusTrader repository Checked 2026-10-31 |
How our number was measured
One continuous run on one NVIDIA RTX 3090 (24 GB) on 2026-10-05: 10-Year Treasury Note futures, 37-minute bars, 2015-01-01 to 2026-09-25, 114,371 bars per evaluation, fees and slippage included. The strategies had single entry and exit conditions, a simple stop-loss and take-profit, and traded one direction only, with the correlation filter on. One evaluation is one completed backtest of a candidate, and the search can evaluate the same candidate more than once. It counts work done. It is not the number of strategies that passed validation, and it says nothing about live results. Different data, settings and cards give different figures. The full record is on the benchmark page.
How to read a speed claim
Ask four things. What counts as one unit: a bar, a backtest, a strategy, or a pass? How many bars does one test cover? On what hardware? Does the test include costs and order handling, or only signal arithmetic? A figure missing any of these answers tells you little, and that includes a figure with a large number in it.
Frequently asked questions
Which backtesting software is the fastest?
No public test answers that. The platforms below report different units on different hardware, and no independent benchmark we could find runs them all on the same job. The honest way to compare two is to run the same strategy, data and costs on both, on your machine.
Why do so few platforms publish a speed figure?
We can only report what we saw: most pages use words like fast with no measurement behind them. A figure needs hardware, data and a test size to mean anything, and those take work to publish and keep current.
Does a faster backtest give a better strategy?
No. Speed lets you test more candidates, and more candidates raise the odds that the best one looks good by luck. It needs out-of-sample testing to go with it, as the walk-forward article explains.
How to cite this page
Use any of these. To cite one figure, add its number to the address (each figure has its own anchor, such as #s12), and if you can, cite the original source linked beside it as well.
Tools and reading
- Algo trading statistics 2026
- Benchmark record
- Monte Carlo simulator
- Equity curve simulator
- What is backtesting?
- Backtesting software for futures
Research software, not investment advice. No profit guarantee; backtest results do not indicate future performance.