Golddigger

Golddigger is a C++17 machine-learning project for training and running short-horizon XAUUSD price predictors with dlib.

It currently supports:

• offline model training from candle CSV files
• hyperparameter tuning for SVR (C, epsilon-insensitivity, gamma)
• daily patching of rolling Dukascopy candle CSV files
• live prediction from Dukascopy CLI candle data
• live prediction from Alpha Vantage FX_INTRADAY

The project builds four binaries in the project root:

• trainer
• predictor
• tuner.bin
• patcher.bin

What It Does

Golddigger builds an SVR model on top of engineered candle features such as:

• recent log returns
• RSI
• ADX
• relative ATR
• relative SMA20 / SMA50
• SMA spread
• cyclical hour-of-day and day-of-week features

The regression target is a signed best-horizon move measured from the current candle close:

• m15: best move reached within the next 6 candles
• h1: best move reached within the next 4 candles
• d1: best move reached within the next 4 candles

Features are still derived from candle closes. Labels use the future horizon's highest high and lowest low, then keep whichever log-return has the larger absolute magnitude. This makes the model learn entry points that lead into the strongest subsequent expansion rather than only the very next candle close.

The training flow is:

1. tune C, epsilon-insensitivity, and gamma with tuner.bin
2. save the tuned parameters to Models/tuner_<timeframe>.dat
3. run trainer to build a model
4. run predictor to fetch live candles and emit predictions

trainer automatically loads the matching tuner_*.dat file when one exists for the training dataset timeframe.

Dependencies

Build/runtime dependencies used by the current codebase:

• CMake
• a C++17 compiler
• dlib
• nlohmann-json
• libcurl
• Node.js and npx if you want to use the Dukascopy provider
• an Alpha Vantage API key if you want to use the Alpha Vantage provider

This project also depends on local code from your HotBits repo:

• cBaseWorker_V2.h
• cHTTPClient.cpp
• cHTTPClient.h

By default, CMake expects those files under:

../Repository/Software/ThreadComponent

If your checkout lives somewhere else, pass the path explicitly:

cmake -S . -B build -DTHREAD_COMPONENT_DIR=/path/to/Repository/Software/ThreadComponent

Homebrew Notes

On macOS, the current project configuration already looks for Homebrew installs of nlohmann-json. A typical setup is:

brew install nlohmann-json dlib curl

Build

Configure and build:

cmake -S . -B build
cmake --build build

Executables are written to the project root:

./trainer
./predictor
./tuner.bin
./patcher.bin

Data Files

The repo currently includes a sample training file:

Data/xauusd-m15-bid-2024-01-01-2026-03-11.csv

The rolling live-data files patched by the updater use the current naming convention:

Data/xauusd-m15-bid.csv
Data/xauusd-h1-bid.csv
Data/xauusd-d1-bid.csv

CSV format:

timestamp,open,high,low,close,volume

Models and Tuned Parameter Files

Typical output files:

• tuned hyperparameters: Models/tuner_m15.dat, Models/tuner_h1.dat, Models/tuner_d1.dat
• trained models: Models/gold_digger_m15.dat, Models/gold_digger_h1.dat, etc.

If you pass custom output paths, those are used instead.

Binary Usage

tuner.bin

Tunes SVR hyperparameters using dlib::find_max_global and cross_validate_regression_trainer.

Default behavior:

• uses Data/xauusd-m15-bid-2024-01-01-2026-03-11.csv
• writes Models/tuner_m15.dat
• uses 50 max optimizer calls
• tunes on up to 5000 evenly spaced samples using 3 folds by default
• derives the epsilon-insensitivity search range automatically from the label distribution
• derives the gamma search range automatically from normalized sample distances

Usage:

./tuner.bin
./tuner.bin --max-calls 25 Data/xauusd-h1.csv
./tuner.bin --max-calls 50 --progress-seconds 60 Data/xauusd-m15-bid.csv
./tuner.bin --max-calls 25 --tuning-samples 8000 --folds 4 Data/xauusd-h1.csv
./tuner.bin --epsilon-range auto Data/xauusd-m15-bid.csv
./tuner.bin --epsilon-range 0.00005:0.003 Data/xauusd-m15-bid.csv
./tuner.bin --gamma-range auto Data/xauusd-m15-bid.csv
./tuner.bin --gamma-range 0.0001:3 Data/xauusd-m15-bid.csv
./tuner.bin Data/xauusd-h1.csv=Models/tuner_h1.dat
./tuner.bin Data/xauusd-m15.csv=Models/tuner_m15.dat Data/xauusd-h1.csv=Models/tuner_h1.dat

Useful flags:

• --max-calls N
• --progress-seconds N
• --tuning-samples N
• --folds N
• --epsilon-range auto|MIN:MAX
• --gamma-range auto|MIN:MAX

Argument format:

• data.csv Result path is inferred as Models/tuner_<timeframe>.dat
• data.csv=output.dat Uses the explicit output file

Notes:

• tuning jobs run on worker threads, so progress can be printed while cross-validation is still running
• the tuner now keeps solver tolerance fixed and tunes epsilon as the SVR epsilon-insensitivity parameter
• by default, tuning uses an evenly spaced subset of the generated samples so large m15 datasets do not spend hours inside the first cross-validation call
• by default, the tuner derives a sensible epsilon-insensitivity search range from the dataset's return distribution
• by default, the tuner derives a sensible gamma search range from normalized sample distances
• --epsilon-range MIN:MAX lets you override that auto-derived range manually
• --gamma-range MIN:MAX lets you override the auto-derived gamma range manually
• Ctrl+C requests a graceful stop and waits for the current evaluation boundary

trainer

Trains a model from one or more CSV datasets.

Default behavior:

• uses Data/xauusd-m15-bid-2024-01-01-2026-03-11.csv
• writes Models/gold_digger_m15.dat

Usage:

./trainer
./trainer Data/xauusd-h1.csv
./trainer Data/xauusd-h1.csv=Models/gold_digger_h1.dat
./trainer Data/xauusd-m15.csv=Models/gold_digger_m15.dat Data/xauusd-h1.csv=Models/gold_digger_h1.dat

Notes:

• training jobs are launched as worker threads
• duplicate model output paths are rejected
• if Models/tuner_<timeframe>.dat exists, the trainer loads it automatically
• if no tuning file exists, trainer falls back to built-in defaults
• Ctrl+C requests a graceful stop and waits for worker threads to exit

predictor

Loads a trained model, fetches live candles from a provider, keeps track of the next missing candle timestamp, and predicts the signed best move over the configured future horizon after each newly completed candle becomes available.

Supported providers:

• dukascopy
• alphavantage

General usage:

./predictor --model Models/gold_digger_m15.dat --instrument xauusd --timeframe m15

Useful flags:

• --provider dukascopy|alphavantage
• --model PATH
• --instrument SYMBOL
• --timeframe m15|h1|d1
• --poll-seconds N
• --availability-delay-seconds N
• --max-predictions N

Provider-specific flags:

• Dukascopy: --dukascopy-command "npx dukascopy-node"
• Alpha Vantage: --alpha-vantage-api-key KEY --alpha-vantage-base-url URL

Predictor With Dukascopy

The Dukascopy provider shells out to:

npx dukascopy-node

Example:

./predictor \
  --provider dukascopy \
  --model Models/gold_digger_m15.dat \
  --instrument xauusd \
  --timeframe m15 \
  --dukascopy-command "npx dukascopy-node"

Optional environment override:

export DUKASCOPY_NODE_COMMAND="npx dukascopy-node"

Predictor With Alpha Vantage

Example:

export ALPHAVANTAGE_API_KEY=your_api_key

./predictor \
  --provider alphavantage \
  --model Models/gold_digger_m15.dat \
  --instrument xauusd \
  --timeframe m15

You can also pass the key explicitly:

./predictor \
  --provider alphavantage \
  --alpha-vantage-api-key your_api_key \
  --model Models/gold_digger_m15.dat \
  --instrument xauusd \
  --timeframe m15

Current Alpha Vantage limitations:

• implemented against FX_INTRADAY
• supports m15 and h1
• d1 is not supported through the current Alpha Vantage provider
• volume is not provided by the API, so predictor uses 0.0 for volume on that provider

Live Prediction Timing

The predictor intentionally avoids requesting the still-forming candle.

It uses:

• candle timeframe
• an availability delay
• polling

So if a provider usually lags one or more minutes after candle close, use:

./predictor --availability-delay-seconds 120

Default availability delay is 60 seconds.

Daily Data Patching

patcher.bin

Fetches one or more completed UTC trading days of Dukascopy candles and merges them into the rolling CSV files in Data/.

Default behavior:

• uses Dukascopy CLI via npx dukascopy-node
• patches m15, h1, and d1
• targets the previous UTC calendar day
• writes into Data/xauusd-m15-bid.csv, Data/xauusd-h1-bid.csv, and Data/xauusd-d1-bid.csv

Usage:

./patcher.bin
./patcher.bin --date 2026-03-31
./patcher.bin --from 2026-04-01 --to 2026-04-06
./patcher.bin --from "2026-03-31 00:00" --to "2026-04-07 23:45" --timeframes m15
./patcher.bin --timeframes m15,h1
./patcher.bin --dukascopy-command "npx dukascopy-node" --data-dir Data

Useful flags:

• --date YYYY-MM-DD
• --from YYYY-MM-DD[ HH:MM[:SS]]
• --to YYYY-MM-DD[ HH:MM[:SS]]
• --instrument SYMBOL
• --data-dir DIR
• --timeframes m15,h1,d1
• --dukascopy-command CMD

Notes:

• requests are formatted in UTC/GMT for Dukascopy
• duplicate timestamps are avoided when patching
• if a timestamp already exists in a CSV, the downloaded candle replaces that row
• --from and --to patch an inclusive UTC date range with one Dukascopy request per timeframe across the whole window, which is useful for catching up missed days without tripping over weekends or holidays
• --from and --to also accept exact UTC datetimes; when a time is included, --to is treated as the last candle timestamp you want included
• date flags accept both YYYY-MM-DD / DD-MM-YYYY and YYYY-MM-DD HH:MM[:SS] / DD-MM-YYYY HH:MM[:SS]
• this is designed as a one-shot updater, so you can schedule it externally for 01:00 GMT

Typical Workflow

1. Tune hyperparameters

./tuner.bin --max-calls 50 Data/xauusd-m15-bid-2024-01-01-2026-03-11.csv

2. Train a model

./trainer Data/xauusd-m15-bid-2024-01-01-2026-03-11.csv=Models/gold_digger_m15.dat

3. Run live prediction

./predictor \
  --provider dukascopy \
  --model Models/gold_digger_m15.dat \
  --instrument xauusd \
  --timeframe m15

4. Patch the rolling candle files

./patcher.bin

Project Layout

.
├── CMakeLists.txt
├── Data/
├── DataUpdate/
├── Indicators/
├── MarketData/
├── Models/
├── Prediction/
├── Training/
├── Utils/
├── patcher.cpp
├── predictor.cpp
├── trainer.cpp
└── tuner.cpp

Notes

• The predictor prints an action signal based on predicted price change vs estimated spread.
• The project uses the current working directory for relative paths like Data/... and Models/....

License

This project is licensed under the GNU General Public License v3.0.

See LICENSE for the full text.

Copyright (C) 2026 Hotbits

Third-party dependencies such as dlib and nlohmann-json remain under their own respective licenses.