MLP Calculator v20 - Backtesting

Network Architecture

Input Neurons (1-10, taken from the data):

Number of Hidden Layers (2-5):

Output Neurons:

Output Activation:

Training Parameters

Learning Rate:

Epochs:

Batch Size:

Random seed: Same seed + same settings = exactly the same result

Shuffle training data order each epoch (recommended)

Loss Function:

💡 Tip: Loading a sample dataset (tab "Data") automatically applies the recommended settings and creates the network. Sigmoid suits binary classification, ReLU deep networks, Linear as output for regression. Start with low learning rates (0.01-0.001) for stable convergence.

Sample Datasets

Loading applies the recommended settings (document "Recommendations for Training Parameters") and creates the network.

Your Own CSV File

Upload CSV file:

Choose File No file chosen

ℹ️ Notes on CSV upload (location & format)

Location: The CSV file must be stored locally on your computer (or on a drive/USB stick connected to it). The file dialog always opens your own computer – the calculator cannot access files on a website or server directly. Please download online files first.
Privacy: The file is processed only in your browser and is not uploaded anywhere.
Structure: one row per observation, header row with column names recommended (e.g. USD,EUR,CNY,GBP,CHF). By default the last column is the target; after loading, any column can be chosen as target. At most 10 input columns.
Values: original values (not normalized). Columns with text or dates are ignored automatically; rows with missing values are skipped.
Order: for time series in ascending chronological order (oldest first) – the last rows then form the test set, as in a real forecast.
Separator: comma or semicolon; with semicolons a decimal comma is allowed (e.g. German Excel export). Encoding UTF-8.

Data Preparation

Changes apply immediately to the loaded data – no reload needed. Retrain afterwards.

Normalize data to [0, 1] (recommended for large or differently scaled values; input and output stay in original units)

Use the target's previous-day value as an additional input (time series only; the first row is dropped)

Enable train/test split (recommended!)

Training share: % (rest = test data for backtesting) Split mode:

Data Overview

Start Training

Current Epoch

Training Loss

0.000

Best Loss

∞

Training Time

Training Log

Live Loss Curve

🎯 Backtesting: The trained network is applied to the training data and to the test data not used in training. All values in original units. The test metrics are what tell you the forecast quality.

Network Architecture

💡 Tip: The network shows the connections between the neurons

Active Neurons (Activation Status)

Shown: ● Normal ● Weak (faded) ● Strong Inactive neurons: grey dot only

Weight Distribution

Activation Values (Table)

💡 Shows activation values of all neurons after the last prediction.
Make a prediction, then click "Update".

Manual Input (original values)

Batch Predictions

Upload CSV file:

Choose File No file chosen

Same columns as in training (original values). If the file also contains the target column, errors and metrics are calculated. Format notes as in tab "Data".

Batch from the test data of the loaded dataset (not used in training; for the currency examples the last period):

Network Statistics

Layer Information

Model Data

Training & Predictions

Visualizations

📝 Note: All data is exported as CSV (Excel-compatible). Graphics as PNG with white background.

🧠 Multi-Layer Perceptron Calculator

Network Architecture

Training Parameters

Sample Datasets

Your Own CSV File

Data Preparation

Data Overview

Start Training

Current Epoch

Training Loss

Best Loss

Training Time

Training Log

Live Loss Curve

Network Architecture

Active Neurons (Activation Status)

Weight Distribution

Activation Values (Table)

Manual Input (original values)

Batch Predictions

Network Statistics

Layer Information

Model Data

Training & Predictions

Visualizations