Build trading systems with automatic data scaling, 3 scaler families (CPU + GPU), and GPU support via OpenCL — all from MQL5.

---

Features

3 scaler families, CPU + GPU, float & double

// CPU (template<TFlt>, TFlt = float or double)
TSN::MaxMinScaler<TFlt>          scaler;
TSN::StandardizationScaler<TFlt> scaler;
TSN::RobustScaler<TFlt>          scaler;

// GPU via OpenCL (same interface, same TFlt)
TSN::MaxMinScalerCL<TFlt>          scalerGPU;
TSN::StandardizationScalerCL<TFlt> scalerGPU;
TSN::RobustScalerCL<TFlt>          scalerGPU;

Every scaler shares the same API: Fit, Transform, InverseTransform, FitTransform, plus Save / LoadByFile / LoadByArr to persist and reload a trained scaler. All CPU and GPU implementations are cross-checked to produce equivalent results.

Matrix scaling

void OnStart()
{
  matrixf mtx;
  mtx.Resize(16000, 16000);
  mtx.Random(0, 20.0);

  TSN::MaxMinScaler<float> scaler;
  matrixf out;
  scaler.FitTransform(mtx, out);
}

Custom column ranges & in-place reconstruction

Scale only a slice of your columns and choose whether the output keeps the full matrix shape (untouched columns preserved) or returns only the scaled slice:

TSN::MaxMinScaler<double> scaler;
scaler.SetRangeEscaler(2, 5);     // scale 5 columns starting at index 2
scaler.SetReconstruir(true);      // output keeps original shape/columns
scaler.FitTransform(mtx, out);

Error handling

Every operation returns bool and sets a descriptive error you can query instead of failing silently:

if(!scaler.FitTransform(mtx, out))
  Print("Scaler error: ", EnumToString(scaler.GetLastScalerError()));

GPU version via OpenCL (real benchmarks)

TSN::MaxMinScalerCL<float> scalerGPU;
scalerGPU.Fit(mtx);
scalerGPU.Transform(mtx, out);   // GPU

TSN::MaxMinScaler<float> scalerCPU;
scalerCPU.Fit(mtx);
scalerCPU.Transform(mtx, out);   // CPU

GPU = 1,268,040 us
CPU = 2,344,502 us

Benchmark on a 9,600 x 9,600 matrix using an Intel UHD Graphics 620 GPU. See Src/Scaler/Bench/README.md for the full set of internal benchmarks (pattern selection for min/max, mean, matrix copy strategy, robust scaling, and the OpenCL upload strategy) that drove the current implementation's performance choices. Used GPU: device #0: GPU 'Intel UHD Graphics 620' with OpenCL 3.0 (24 units, 1100 MHz, 3240 Mb, version 30.0.100.9670, rating 836)

---

Quick Start

1. Install repository

See Instalation section.

2. Include the scaler in your script

void OnStart()
{
  matrixf data;
  data.Resize(100, 10);
  data.Random(0, 100.0);

  TSN::MaxMinScaler<float> scaler;
  matrixf out;
  scaler.FitTransform(data, out);

  Print("Scaling complete. Sample value: ", out[0][0]);
}

3. Verify the installation

Run any script from the Test/ folder in the MetaTrader 5 Editor. If no errors appear in the log, you're all set.

---

Repository Structure

ScalerByLeo/
├── Src/
│   ├── Scaler/
│   │   ├── MinMax/   # MaxMinScaler (CPU) + MaxMinScalerCL (GPU)
│   │   ├── Std/      # StandardizationScaler (CPU) + StandardizationScalerCL (GPU)
│   │   ├── RScaler/  # RobustScaler (CPU) + RobustScalerCL (GPU)
│   │   ├── Bench/    # Internal micro-benchmarks used to pick the fastest patterns
│   │   ├── AllImp.mqh  # Single include for all CPU scalers
│   │   └── AllCL.mqh   # Single include for all GPU scalers
│   ├── CL/       # OpenCL context/kernel manager
│   └── Kernels/  # OpenCL kernel sources
├── Test/       # Equality tests and benchmarks (CPU vs GPU)
└── Images/     # Repository banner

---

License

Read Full License

By downloading or using this repository, you accept the license terms.

---

Requirements

See the dependencies.json file.

---

Installation of repo code

cd "C:\Users\YOUR USER\AppData\Roaming\MetaQuotes\Terminal\YOUR ID\MQL5\Shared Projects"
tsndep install "https://forge.mql5.io/nique_372/MQLArticles.git"

• Using the tsndep command requires the tsndep package (available on PyPI). This command automatically downloads all dependencies and installs every requirement from the repositories.
• If any part of the system is private, the install will fail. Contact me so I can grant you access (if it's a product, you can purchase it; if you have any questions, don't hesitate to reach out). Check the dependencies.json file for more info.

---

Contact

• Platform: MQL5 Community
• Profile: https://www.mql5.com/es/users/nique_372/news

---

Copyright © 2026 Nique-Leo.