Describe a slow run
For quants · algo-trading developers · platforms · funds

Faster backtests, parameter sweeps and pricing in Python. The same results, bit for bit.

TAKT makes the native code under quant workloads run faster and keeps every output identical to the original. Three libraries are already available as free drop-in packages for Python. For your own workload, we measure first and agree the price and acceptance before work starts.

“Bit for bit” means that every output value is compared with the original program’s output as raw bytes, on the same machine, not within a tolerance. Other terms are explained at the end of the page.

Path 1

Individual quants and algo developers

Three free packages on PyPI keep the official Python package and replace the native library inside it. Your code still says import talib, import QuantLib or import lightgbm. Only the library underneath changes. For now they are built for Linux x86-64 only.

PackageReplacesFaster in the free packageRuns the original code at the original speed
ta-lib-takt 0.8.2TA-Lib 0.8.2 (Python 3.11+; 3.9 and 3.10 get 0.8.1) CCI, SAR, DX, ADX, EMA, STOCH and RSI on long series: 2.0–6.0× in a parameter sweep over one million bars Short series (a few thousand bars), all other functions, CPUs without AVX2 and FMA
quantlib-takt 1.43QuantLib 1.43 European option Monte Carlo, including Sobol and Greeks by bumping: a book of options about 5–8× faster, Greeks up to 18× Asian, American, barrier, basket, Heston and local volatility engines (these are accelerated in the full version)
lightgbm-takt 4.7.0.post2LightGBM 4.7.0 Permutation importance, partial dependence and ICE, MDA and what-if sweeps on models up to 300 trees: up to 3.2× for the whole process Training, ordinary batch prediction, SHAP, larger models (the full version accelerates larger models and batch prediction)

Check it on a public workload, on your machine

Install the official package and ours into two separate environments, run the same script in both, and compare the output and the time. Your code and data stay on your machine; nothing is sent to TAKT.

python3 -m venv official && official/bin/pip install "TA-Lib==0.8.2"   # on Python 3.9/3.10: "TA-Lib==0.8.1"
python3 -m venv takt && takt/bin/pip install ta-lib-takt
curl -fLO https://taktcycles.com/showcase/talib/tools/py_sweep.py
time official/bin/python py_sweep.py CCI   # prints a digest of the results
time takt/bin/python py_sweep.py CCI       # the same digest, in less time

py_sweep.py is the published sweep from our TA-Lib measurements: 99 periods of one indicator over one million generated bars. Then run a script of your own in both environments, for example a public strategy from your framework. Save the arrays with numpy.save and compare them with numpy.array_equal(a, b, equal_nan=True). Compare QuantLib against QuantLib==1.43 and LightGBM against lightgbm==4.7.0. In an existing environment, uninstall the official package first, because both provide the same import.

Gains depend on the function, the series length and the data, so measure your own case. Prefer not to install anything into Python? The EMA demo runs the published native builds on your machine with one script, without an account.

Strategy optimizers: vn.py, freqtrade, backtrader

  • vn.py. Download our program (MIT, Linux x86-64) and run it next to vn.py’s own optimizer on the same generated data. It prints the same ten best parameter sets as vn.py. What was measured.
  • freqtrade hyperopt and backtrader optstrategy. Both projects are GPL-3.0, so we do not distribute these builds. Instead, we run your hyperopt or optimization on our machines and return the same result file. The workload you send can be a public strategy, for example freqtrade’s SampleStrategy, so you can see the result and the time before sharing anything of your own. The first 20 core-hours are free. After that, it costs $0.10 per core-hour of our machine time, paid from your dashboard balance.
  • Where ta-lib-takt does not help. A freqtrade hyperopt computes its indicators once per run. In our measurement that took 0.6 s of a 141 s run, so a faster TA-Lib alone does not shorten a hyperopt. The freqtrade numbers below come from the separate build that we run for you.
Path 2

Trading platforms and tool vendors

Does your product ship a backtester, an indicator library, a pricing engine or a model-analysis tool? Then a TAKT build can replace its slow part and keep your API. Your users’ code and your Python bindings stay the same, and the output stays identical.

  • What you getEither a drop-in library with the same ABI, as with TA-Lib, QuantLib and LightGBM above, or a native build of a Python hot loop with the same output, as with vn.py. It comes as a signed package with a report of the measurements and the equivalence checks.
  • Measured on your workload firstYou name the workload, the platform and the outputs that must match. C, C++ and Rust projects built with CMake, make or Cargo go through the dashboard, and the speedup is measured before you pay. We call this Level 1: from $290 per project, and nothing to pay below a 10% speedup. Python code and library replacements are individual engineering work, called Level 2. A written order sets the price, timetable, guaranteed minimum, acceptance procedure and refund table before anything is charged: 30% upfront, the rest after acceptance.
  • LicenceThe licence is perpetual, worldwide and non-exclusive. You may run the build, integrate it into your products and services, and distribute it in binary form as part of them. The build contains no licence check, expiry date, remote switch or network call, and it keeps working without any TAKT service. Standalone resale of the build is not included. Terms of Service, clause 8.
  • New releases of your codeA CI subscription optimizes and checks every new release again. It costs from $149 a month, billed yearly. See the pricing.
  • Full versions of the published librariesThe full QuantLib build accelerates every Monte Carlo engine listed above. The full LightGBM build also accelerates larger models, batch prediction and SHAP’s permutation explainer. Both add the C/C++ library and support, from $399 a month, billed yearly.
  • ContinuityThe source of your build can be put into escrow before payment. See source escrow below.
Path 3

Funds and prop shops

Our published terms already cover the points below. The full texts are linked, and they govern.

  • Start without codeDescribe one slow run: how long it takes, how often it runs and on what hardware. Or measure it yourself with the free, open-source takt-harness (MIT) and send us the readings. They contain timings, hashes and machine facts, but no code and no output. We reply with a preliminary range.
  • On-premises pilot: in development, by arrangementIf your code and data must stay inside your infrastructure, we can discuss a pilot for your own servers or private cloud account. The complete on-premises optimization environment is not yet ready for deployment. A signed order would set the development and qualification timetable, scope, price, acceptance and refunds before any installation or charge.
    • The proposed service would process your materials and deliver builds, reports and the measurement passport inside the agreed perimeter. The environment must pass the agreed qualification tests before it receives your materials.
    • The environment would initiate no network connection and could run with no route outside. You would choose whether to send the limited records specified in the order, after inspecting them. These may include hashes, measurements and platform evidence, not source text, input or output contents, or binaries; they may still contain confidential or personal information.
    • Staff access would be through sessions you open and can end. Remote display makes the displayed information available outside your perimeter and requires separate approval of the channel and recipients. If your policy prohibits that, support must be on site or use synthetic materials. TAKT would not send your materials to an AI provider.
    • The build would be measured on the hardware and metric agreed for your perimeter. No speed-up or platform support is promised before it has been measured or qualified.
    • Qualification would use synthetic materials and assess the specific platform, configuration, controls and trust assumptions. A confidential-computing label alone is insufficient and does not guarantee protection against every host or physical attack. If qualification fails, no operational deployment proceeds; only an expressly agreed qualification fee remains payable and other prepaid fees are returned.
    • The intended initial scope is automatic optimization of C, C++ and Rust (Level 1), subject to the signed order and qualification. Python and other engineering work are outside that scope unless separately agreed under the same data-handling restrictions. Pricing is by quotation.
    Terms of Service, clause 5.11, and the On-Premises Rider set the proposed contractual framework.
  • Or send only the hot partFor a build on our servers we need the code that runs in the hot loop, and only that. Usually this is the simulator, the indicators and the portfolio state, without the business logic around them. We help you extract it.
  • For work on our servers: confidentiality and controlled executionThe NDA applies before upload. Files are encrypted in your browser before transmission. Automatic platform builds and checks run in single-use VMs without network access on our own server in Lithuania, with a separate project storage key. Level 2 engineering may additionally use a separate workspace outside those VMs, subject to verified isolation; its compilation and execution remain without network access. The proposed AI-assisted route uses local coding agents with commercial Anthropic and OpenAI APIs under TAKT's supervision, only after the required controls are ready and you give separate permission for that project and order. It may send source code, open inputs and relevant outputs to those services, subject to the exclusions in Terms 5.12. You can withdraw permission; requests already sent cannot be recalled and provider copies have separate retention rules. Your materials are not used for other clients, and Level 1 and automatic checks do not send them to external AI services. An AI-free Level 2 alternative is not a standing offer. See the operating limits, Terms 5.12 and the Recipients Schedule.
  • Hidden inputsPart of your data is held back until the candidate build is fixed, so the speedup cannot be fitted to that test. For work on our servers, you send it encrypted and supply the key when the build is ready. For an on-premises order, that procedure takes place inside your perimeter under the agreed acceptance protocol.
  • Acceptance on agreed hardwareBefore work starts, a measurement passport records the machine, the workload, the metric, the threshold and which outputs must be identical. The result is judged against that record. For the standard hosted offer, a Level 1 build can first be tried on your own hardware under the Trial Agreement. An on-premises pilot uses the acceptance and payment procedure in its signed order; the standard trial does not apply automatically. For individual work, the acceptance procedure is part of the written order. The Enterprise CI plan (from $1,500 a month, billed yearly) includes acceptance on your hardware.
  • Check the build yourselfEvery package is signed and comes with:
    • an SBOM;
    • a script that shows, on your runs, that the build makes no network calls, starts no programs and writes no files;
    • a byte-for-byte comparison with your original build on your inputs;
    • a shadow mode that runs both builds side by side in a test environment.
    You may security-test the build, including through contractors bound by confidentiality. Trust and delivery.
  • Correctness commitmentSuppose that within twelve months you report a reproducible difference from your original code, on an input within the agreed domain. Then we deliver a corrected build within 30 days or refund the price of that build.
  • Source escrow, if you need itFor the standard hosted offer, source escrow is available on written request before payment under Terms 8. For an on-premises order, escrow is included only if the signed order specifies a compliant arrangement preserving that order's data-location and disclosure commitments; the hosted provisioning procedure does not apply automatically. This is a sealed bootable image that you run, at your cost, on the cloud platform recorded for the order. It gives out the source only on the release events listed in the escrow box rider. If the platform is not yet qualified, the escrow order is not confirmed or paid.
  • ProcurementPublished documents: Terms, NDA, DPA, Privacy Notice and Recipients Schedule. Security questionnaire answers, the data-flow description and an SBOM sample are available on request from [email protected]. We do not yet hold SOC 2 or ISO 27001 certification.
  • Live tradingWe speed up a live path only if profiling shows that computation is the bottleneck. Then we measure the whole path from the market event to the order. Our strongest results are in batch work: backtests, training and parameter sweeps.
Published measurements

Measured on public code, same output as the original

Each case compares the unmodified original with the TAKT build on the same machine, on AMD Zen 3 and Zen 4 under Linux x86-64. The scripts and readings are published. These are results on specific code, not a forecast for yours.

16–20×vs a competent native port

vn.py optimizer

vn.py’s brute-force optimizer over 600 parameter sets of its stock moving-average strategy. We get the same top 10 with the same Sharpe ratios, returns, drawdowns and trade counts. It runs about 12,000× faster than vn.py itself, mostly because it leaves Python.

6–11×whole hyperopt run, same result file

freqtrade hyperopt

freqtrade’s own hyperopt on five crypto pairs and two years of 5-minute candles: 5.9–6.7× over 30 epochs and 7.2–11× over 100 epochs. Every epoch and every trade in the result file is identical. 15 cases and 204,132 trades were compared. Available as a hosted run.

2.2–5.2×whole optimization run, same results

backtrader optstrategy

The SmaCross grid from backtrader’s documentation runs 4.7–5.2× faster, and five other workloads run 2.2–3.9× faster. Every analyzer output, order and trade notification is identical. 25 cases and 91,022 trades were compared. Available as a hosted run.

2.0–6.0×from Python, wrapper unchanged

TA-Lib 0.8.1

A drop-in C library with the same ABI. 4.59 million calls were checked against the official library with no difference. Seven indicators run 2.3–7.4× faster from C in a sweep over one million bars. Free as ta-lib-takt.

5–18×European MC, free package

QuantLib 1.43 Monte Carlo

Through the unchanged Python API, a book of European options prices about 5–8× faster, and Greeks by bumping up to 18× faster. In the full version, the Asian, American, barrier, basket, Heston and local volatility engines run 2–9× faster. Every result is identical to the official package.

up to 3.2×free analyses, default threads

LightGBM 4.7.0

Permutation importance, partial dependence and ICE, MDA and what-if sweeps through the unchanged API, with predictions identical to the official package. The full version reaches up to 19× on one core with a 1,000-tree model.

TAKT is not affiliated with, endorsed or sanctioned by the TA-Lib, QuantLib, LightGBM, vn.py, freqtrade or backtrader projects. Their names only identify the software we measured.

Plain words

Terms used on this page

Bit for bit
Every output value is compared with the original program’s output as raw bytes, on the same machine. Nothing is accepted “within a tolerance”.
Measurement passport
A short written record, fixed before work starts, of the machine, the workload, the tool versions, the metric, the threshold and the outputs that must be identical. The result is judged against it.
Level 1
Automatic optimization of a C, C++ or Rust project through the dashboard. The speedup is measured before you pay.
Level 2
Separately scoped engineering for Python code and library replacements, which may use AI coding agents under TAKT supervision only with the project's permission and the controls described above. The scope, price, guaranteed minimum and acceptance are agreed in a written order first.
Hidden inputs
Part of your test data that you keep encrypted until the build is ready, so the speedup cannot be fitted to the test.
Hosted run
We run your workload, for example a hyperopt, with the faster build on our machines and return the output and the run time.
Core-hour
One hour of one CPU core of our machine.
Who we are

Company and contact

TAKT is the trading name of ABX DEVELOPMENT LLP, registered in Scotland, no. SO301872, 39/5 Granton Crescent, Edinburgh EH5 1BN, United Kingdom.

Design partners: 50% of your first order comes back to your balance as credit. 10 places.