FinGAN-TF

A TensorFlow re-implementation of FinGAN (Vuletic, Prenzel, Cucuringu) with a deterministic LSTM-Fin baseline. The pipeline trains conditional return-forecasting models over US equities and sector ETFs, then evaluates them with the financial loss combinations and trading-style metrics from the original paper.

Reference PyTorch implementation: https://github.com/milenavuletic/Fin-GAN

Layout

src/
├── main.py                         # CLI entry point
├── models/
│   ├── generator.py                # LSTM + dense head, noise + condition → next return
│   ├── discriminator.py            # PnL/MSE/SR/STD/BCE-aware discriminator
│   ├── LSTM.py                     # deterministic LSTM-Fin baseline
│   └── utils.py
├── libraries/
│   ├── data/
│   │   ├── loader.py               # Nasdaq Data Link fetch + ArcticDB cache
│   │   └── preprocessing.py        # excess-return construction, train/val/test split
│   └── training/
│       ├── fin_timegan_trainer.py   # per-ticker FinGAN training over loss combos
│       ├── fin_timegan_universal.py # shared model over a ticker pool
│       ├── lstm_fin_trainer.py      # deterministic LSTM-Fin pipeline
│       └── plots.py
├── utils/
│   ├── evaluation_fintimegan.py    # Sharpe ratio, hit rate, correlation, sanity checks
│   └── lstm_eval.py
└── database/arctic/                # local ArcticDB cache (created on first run)

Run outputs (models, plots, results CSVs, run notes) land under src/runs/<label>/.

Data

Prices are pulled from Nasdaq Data Link (SHARADAR/SEP for stocks, SHARADAR/SFP for ETFs) on first use and cached in a local ArcticDB at src/database/arctic/.

You must supply your own Nasdaq Data Link API key. Open src/libraries/data/loader.py and replace the value of ndl.ApiConfig.api_key with your own key before running the pipeline. Without a valid key, the first fetch for any uncached ticker will fail.

Models

• FinGAN — conditional GAN. Generator: noise + length-l condition window → next return via an LSTM and two dense layers. Discriminator is trained against any combination of PnL, MSE, Sharpe-ratio, std, and BCE objectives.
• LSTM-Fin — deterministic baseline mirroring Vuletic's LSTM class. Same condition window, no noise input.

Loss combinations evaluated by default (one model is trained per combination):

PnL MSE STD,  PnL,  PnL MSE,  PnL MSE SR,  PnL SR,
PnL STD,     SR,   SR MSE,   MSE,         BCE

Usage

From the src/ directory:

# Per-ticker FinGAN over multiple seeds
python main.py --model fingan --train_mode single \
  --tickers AMZN HD KO --seeds 10 20 40

# Universal FinGAN over a preset ticker pool
python main.py --model fingan --train_mode universal \
  --experiment xlp --seed 42

# Deterministic LSTM-Fin baseline
python main.py --model lstm --tickers AMZN HD KO --seed 42

Universal experiment presets (--experiment): full, xlp, xlp_etf.

Common flags

flag	default	meaning
--l	10	condition window length
--pred	1	prediction horizon
--tr / --vl / --h	0.8 / 0.1 / 1	train/val split, holdout
--hidden_dim / --noise_dim	8 / 8	network widths
--n_epochs / --ngrad	100 / 25	training / GradientCheck epochs
--batch_size / --lr	100 / 1e-4	optimisation
--diter / --tanh_coeff	1 / 100	discriminator iters, soft-sign coefficient
--eval_samples	1000	MC samples at evaluation time
--jit_compile	off	enable XLA
--plot / --no-plot	on	save per-run PNGs
--notes_file	auto	path for run_notes.txt

Outputs

Each run writes under src/runs/<label>/:

• models/ — trained weights per loss combo
• plots/ — overlays, correlation heatmaps, gradient-norm and intraday/overnight charts
• results/ — per-seed and aggregated metric CSVs (Sharpe ratio, hit rate, correlation, sanity-check shuffles)
• pnl/ — PnL series
• run_notes.txt — config + summary snapshot

Dependencies

tensorflow, numpy, pandas, matplotlib, tqdm, nasdaqdatalink, arcticdb.