xLSTM-Timeseries

See src/xLSTM for code. Calculates the next-day-ahead price for an asset. Use in conjunction with strong fundamentals and technical analysis to judge when to trade.

Running order

• multi_retrieve_ib_dataset.py - builds historical one-day price datasets based on ticker names e.g. NVDA, GOOGL etc.
• prep_train_xLSTM.py - trains the xLSTM model using historical price data. Generates train, validate, and test data sets.
• predict_xLSTM.py - given the trained model and test data, retrieves the nest day price for each company.

xLSTM-TS model methodology

Methods

• Data source & schema

  • Input file: STABLE_PRICES.txt (multi-ticker table with columns: Date, Ticker, OHLCV; target is Close).
  • Flexible CSV/TXT parsing with automatic delimiter detection; Date coerced to datetime and sorted.

• Business-day alignment & cleaning

  • Per-ticker time series aligned to a business-day index.
  • Missing values handled by time interpolation → forward-fill → back-fill.
  • All infinities and all-NaN rows removed before alignment.

• Leak-safe split & denoising

  • Temporal split per ticker: Train 86%, Validation 7%, Test (remainder).
  • Wavelet denoising (Daubechies-4, soft threshold) applied only to train & validation segments; test remains raw.
  • Denoising is split-local to prevent look-ahead leakage.

• Scaling (per ticker)

  • Darts Scaler Min–Max ([0,1]) fit on denoised train only.
  • The same scaler is applied to denoised validation and raw test segments.
  • Per-ticker train-denoised min/max for Close recorded in artifacts/scaling_params.csv.

• Windowing (supervised samples)

  • Univariate modeling of Close only: inputs (X) are shape ((N, \text{seq_len}, 1)).
  • Horizon H=1 (next-day close).
  • Windows built per ticker within each split; counts reported in artifacts/window_counts_seqLEN.csv.

• Model: xLSTM-TS block

  • Input linear embedding → mLSTM block → sLSTM block → mLSTM block → layer norm + linear head.
  • Each block stacks: LayerNorm → causal depthwise Conv1d (kernels: m=4, s=2) → Multi-Head Attention (2 heads) → LSTM → projection/FF sublayer → LayerNorm with residuals.
  • Tunable embedding dimension; default shown as 64.

• Training objective & optimization

  • Loss: MSE (regression on scaled target).
  • Optimizer: Adam; learning rate scheduled by ReduceLROnPlateau (factor 0.5, patience 10, floor (1e{-8})).
  • Gradient clipping: max-norm 1.0.
  • Early stopping via validation loss with patience 30, up to 200 epochs.
  • Determinism: global seed (42); GPU used if available.

• Hyperparameter search (Optuna TPE)

  • Search space:

    • seq_len ∈ {60, 100, 150, 200, 256}
    • batch_size ∈ {8, 16, 32, 64}
    • embed_dim ∈ {32, 64, 128, 256, 384}
    • lr ~ log-uniform in ([1e{-5}, 3e{-3}])

  • Sampler: TPESampler (multivariate, seeded).

  • Best trial’s state dict and learning curve saved; full trials exported to artifacts/study_trials.csv.

• Evaluation & baselines

  • Splits evaluated with DataLoaders (no shuffling except train).

  • Naïve persistence baseline: ( \hat{y}t = x{t-1} ) (last observed close).

  • Metrics (overall + per ticker): MSE, RMSE, MAE, MAPE, sMAPE, (R^2).

  • Diebold–Mariano test (squared-error loss, (h{=}1)) with Newey–West variance for model vs naïve.

  • Predictions & errors exported for validation and test:

    • artifacts/predictions_val.csv, artifacts/predictions_test.csv
    • artifacts/metrics_val_overall.json, artifacts/metrics_test_overall.json
    • artifacts/metrics_val_by_ticker.csv, artifacts/metrics_test_by_ticker.csv

• Artifacts & reproducibility

  • Preprocessing report: artifacts/data_config.json, artifacts/split_report.csv, artifacts/scaling_params.csv.
  • Scaled splits: scaled_train_businessB_denoised.csv, scaled_val_businessB_denoised.csv, scaled_test_businessB.csv.
  • Best model: artifacts/xlstm_ts_best_state_dict_STABLE_PRICES.pt and artifacts/xlstm_ts_best_hparams_STABLE_PRICES.json.
  • Learning curve for best trial: artifacts/learning_curve_best.csv.

• Runtime details

  • Batch loading via PyTorch DataLoader (train shuffled; val/test not shuffled).
  • Device-agnostic execution (CPU/GPU).
  • Verbose logging toggle with VERBOSE=True.

SilentWealth

Draft project. Day trades bitcoin using an Interactive Broker account and the Interactive Broker Gateway tool.

Arguments

• --ticker_name the trade ticker name e.g. BP., LLOY, SOXS, SOXL, BTC. Always required.
• --exchange the name of the exchange, which also dictates the trading hours. Options are, LSE, ARCA, and NASDAQ (not applicable for BTC).
• --quantity the number of stocks to trade (not applicable for BTC).
• --frame_size trade windows in minutes. Typically, set this to 1 or 5. Always required.
• --account Set to paper or live. Always required.
• --stop_loss 0 to 1 as a fraction from the stock value e.g., 0.01 sets a -1% stop loss below the stock value. Set to 0 to switch off.
• --take_profit a take profit set from the buy position as a percentage of the investment. Default is 0.02 e.g., 2% of the investment value. Set to 0 to switch off.
• --limit_order When set the code tries to make a buy at the lower quartile of the bid-ask range using a limit order. At the moment this does not support stop loss or take profit (do not set either if you wish to place a Limit Order). Warning: If the code fails to fill the order at the requested price within 4 seconds, a market order is made..

Adjusting BUY/SELL conditions:

• --ema_short short-range exponential moving average (default: 9). Optional.
• --ema_medium medium-range exponential moving average (default: 20). Optional.
• --ema_long long-range exponential moving average (default: 200). Set to 0 to exclude from BUY conditions.
• --vwap volume-weighted average price (default: 9). Set to 0 to exclude from BUY conditions.
• --rsi_period units to compute the RSI (default: 14). Set to 0 to exclude from BUY conditions.
• --anchor_distance Pins a buy order within a distance from the short crossing up and over the medium. The longer this value, the more likely the BUY signal will happen shortly before a SELL signal.

Examples

Use the following minimal examples and options provided above to execute the software.

London Stock Exchange

Attempting to make a limit order with default BUY conditions

--ticker_name LLOY --exchange LSE --quantity 100 --frame_size 1 --account paper --limit_order

Attempting to make a limit order with default BUY conditions and anchored within 4 events of the BUY signal

--ticker_name LLOY --exchange LSE --quantity 100 --frame_size 1 --account paper --limit_order --anchor_distance 4

Turing off RSI, VWAP, and long EMA BUY conditions

--ticker_name LLOY --exchange LSE --quantity 100 --frame_size 1 --account paper --limit_order --vwap 0 --rsi_period 0 --ema_long 0

NASDAQ

--ticker_name PLTR --exchange NASDAQ --quantity 100 --frame_size 1 --account paper --limit_order

Bitcoin

--ticker_name BTC --frame_size 1 --account paper --stop_loss_percent 0.01 --dollar_amount 1000

ARCA ETF

--ticker_name SOXS --exchange ARCA --quantity 100 --frame_size 1 --account paper --limit_order

Benchmarking

BTC

1 min windows over 10 days

• Short: 12

• Medium: 30

• Long: 100

• RSI duration: 14

• RSI top: 80

• RSI bottom: 40

• VWAP: True

• ATR duration: 12

• Initial investment: $1000

• Commission pot: $200

• TOTAL PROFIT/LOSS: $

• COMMISSION: $

• NET PROFIT AFTER COMMISSION: $59.724179172215855