← back to the showcase

Strategy and parameter grids: measured time for the same requested result

These case studies measure complete grids, hyperopt jobs and numerical workloads against the original engine or a named reference. The result contract and timing unit matter as much as the factor: an indicator call, a whole backtest grid and a pricing book are different jobs.

Measured results

Elapsed factors below are ratios of median whole-process wall times within each named comparison. CPU-cycle factors are labelled separately. Each linked study retains its declared complete-process boundary: startup, input loading or generation, the requested computation and output, including any in-window network metadata requests. Preparation outside that process is excluded only as defined in the individual study. Each link retains its readings, platform details and conditions.

Measured task Recorded result Baseline and finite workload
vn.py DoubleMaStrategy grid 16.3× elapsed time The study's B2 compiled reference; 600 parameter sets, 200,000 synthetic minute bars, four threads on AMD Zen 3. Case details
backtrader optstrategy 2.24–5.23× elapsed time Unchanged backtrader 1.9.78.123; six named workloads on Zen 3/4, one assigned CPU per run. The main SmaCross grid has 100 parameter sets and 26,304 bars. Case details
freqtrade hyperopt 5.91–11.02× elapsed time Unchanged freqtrade 2026.9; SampleStrategy, five spot pairs, 5-minute candles, 30 or 100 epochs, one job on Zen 3/4. Case details
TA-Lib C indicator workloads 2.34–7.40× user-mode CPU cycles on the selected seven indicators TA-Lib 0.8.1, one million bars, named C-driver single/sweep cases on Zen 3/4; fastest supplied baseline for each metric. The CCI sweep on Zen 4 is 7.11× elapsed time. Case details
QuantLib Monte Carlo, selected C++ cases Barrier: 5.70× cycles / 5.61× elapsed; European Greeks: 23.20× cycles / 22.82× elapsed Official QuantLib 1.43 reference build on Zen 3, the named barrier and European-Greeks workloads. These selected cases do not define a range for every Monte Carlo engine. Case details

The vn.py factor uses the named compiled comparator. Both programs in that vn.py comparison used the same four-core allocation. The required result is the ordered top-10 table with parameters, Sharpe, return, drawdown and trade count at the recorded output precision; the 16.3× figure does not describe every vn.py strategy or every statistic of its optimizer.

For backtrader, the range includes all six recorded workloads: the lowest is TestStrategy at 2.24×, and the highest is the main SmaCross grid at 5.23×. The main grid took 128.7 / 27.10 seconds on Zen 4 (4.75×), and 224.0 / 42.79 seconds on Zen 3 (5.23×), original / TAKT. Strategy code, callbacks and execution settings define the scope of each case.

The freqtrade 30-epoch series used seven interleaved rounds. Its Zen 4 job took 243.0 / 41.08 seconds (5.91×), original / TAKT. The 100-epoch series used only two interleaved rounds; its 11.02× Zen 3 result is a median without a meaningful confidence interval. It must not be read as a confidence bound.

TA-Lib's seven selected indicators are EMA, RSI, ADX, DX, SAR, STOCH and CCI. The range above is for user-mode cycles, not elapsed seconds. ATR, MACD, BBANDS, shorter series and Python-wrapper workloads have their own smaller-gain cases in the linked report. In the linked freqtrade experiment, substituting the faster TA-Lib alone did not measurably shorten the complete hyperopt job.

The two QuantLib factors are C++ benchmark results. The barrier case uses 252 steps and 10,000 paths; the European-Greeks case uses 252 steps and 5,000 antithetic paths with six pricings. Python startup, other option engines and other systems have separate results on the existing page.

What matched

Study Required measured output
vn.py The complete ordered top-10 table: parameter values, Sharpe, return, drawdown and trade count, printed at 17 significant digits. The full printed table matched byte for byte.
backtrader The complete canonical result for every parameter set: analyzer outputs, every order/trade notification, final portfolio value, cash and positions. The linked comparison defines the canonical representation.
freqtrade Every epoch's parameters, loss, metrics, trades and orders in the complete canonical result. Exactly two per-epoch wall-clock timestamps are excluded.
TA-Lib C API return status, output bounds and every output value, compared bit for bit under the recorded array checks.
QuantLib The complete recorded price, estimate and Greek output, printed at 17 significant digits and as hexadecimal floats, with byte-identical comparisons on the same system.

The same inputs, grids or requested epochs are used on both sides, with the recorded random seeds and ordering where applicable. Native-library versions and platform maths belong to the comparison. In particular, vn.py's own NumPy-based Sharpe result can differ in its last digits between AVX2 and AVX-512 systems. QuantLib's official result can differ across system maths-library versions. Equality is established against the reference on the stated system, not by assuming that all machines print the same bits.

Scope and current conditions

These factors belong to their recorded fixtures and builds. They do not combine into one universal replacement for trading and pricing engines. An indicator factor and a complete-engine factor are separate measurements and must not be multiplied.

The existing case pages retain their current release checks, coverage, availability and terms. A release check covers its stated inputs and CPU; it does not repeat every earlier measurement campaign. The table above keeps the identity of the recorded campaigns.

Upstream project licences include MIT for vn.py, GPLv3 for backtrader and freqtrade, BSD-3-Clause for the TA-Lib C library and QuantLib, and BSD-2-Clause for the TA-Lib Python wrapper. Third-party strategy and data files retain their own notices and licence conditions. These licence facts do not establish the terms of a different future deployment or redistribution.

Measure one complete workload

Describe the engine and version, strategy or pricing task, parameter grid or epochs, bars and feeds, required result, CPU allocation and run frequency. A representative complete job and its output contract are the basis for a scoped comparison; the fixture factors above do not predict its outcome.

Describe a strategy or parameter-grid workload.

Project names identify the measured software; no endorsement is claimed. Trading profit, model quality and a customer's cost saving are not measured by these runtime comparisons.

Source text: README.md

Telegram