PutStrike iTransformer — Per-Stock Forecasting Models

iTransformer (ICLR 2024) — individual models trained per stock for 60-day price forecasting.

Per-Stock Models

{len(per_stock_metrics)} individual per-stock models, each under per_stock/{SYMBOL}.onnx.

• Architecture: iTransformer with RevIN (533,141 parameters each)
• Input: 60 days x 135 features (OHLCV technicals + macro + relative strength)
• Output: 60-day forward return forecast
• Training: Walk-forward validation (70/15/15), HuberLoss(delta=0.02)
• Average 7-day directional accuracy: 51.3%
• Average 30-day directional accuracy: 56.0%

v11.0 Features (172 total — includes gamma squeeze + sentiment + stock-specific drivers + tail risk + style rotation)

• 83 original: OHLCV technicals, momentum, volume, volatility, statistics, calendar, macro
• 4 relative strength: stock vs SPY returns (5/20/60d), SPY correlation
• 3 cross-asset: VIX correlation, rolling beta, volume-price correlation
• 4 advanced volume: MFI, A/D line, VWAP deviation, Force Index
• 5 price structure: range position, ATR ratio, consecutive days, candle body
• 5 statistical regime: Hurst exponent, Parkinson/GK volatility, consistency, tail ratio
• 4 intermarket: SPY momentum, gold/oil ratio, DXY-VIX interaction
• 3 sector ETF: stock vs sector ETF returns (5/20d), sector correlation
• 4 credit market: HYG/TLT returns, credit spread proxy, HYG-SPY divergence
• 1 VIX term structure: VIX9D/VIX short-term fear ratio
• 2 industry commodity: per-stock commodity correlation and return
• 2 intermarket extended: copper/gold ratio change, BTC sentiment
• 6 gamma squeeze proxies: volume acceleration, price-volume momentum, range expansion, gap acceleration, squeeze breakout signal, volume-price impact
• 4 market breadth & rotation: tech rotation (QQQ-SPY), small cap rotation (IWM-SPY), semiconductor momentum (SOX), biotech momentum (XBI)
• 4 sentiment proxies: realized/implied vol ratio, VIX-SPY short corr, credit momentum 10d, fear composite
• 4 stock-specific drivers: per-company primary/secondary driving asset returns & correlations (90 unique driver mappings across SOX, IGV, HACK, KRE, ITA, XOP, IBB, XHB, XRT, LIT, etc.)
• 2 FRED extended: St. Louis Fed Financial Stress Index, 10Y-3M yield spread
• 2 FRED rates: Federal Funds Rate level and 20d change (monetary policy stance)
• 2 FRED FX: JPY/USD 20d change and z-score (carry trade unwinding proxy)
• 2 tail risk: CBOE SKEW level and 20d z-score (options tail risk pricing)
• 1 value/growth rotation: IWF vs IWD 20d return spread
• 1 risk appetite: XLY vs XLP 20d return spread (consumer discretionary vs staples)

Usage

import onnxruntime as ort
import numpy as np

# Load per-stock model
session = ort.InferenceSession("per_stock/AAPL.onnx")

# features shape: (1, 60, 135)
output = session.run(None, {"features": features})[0]
# output shape: (1, 60) — predicted returns

Disclaimer

This model is for research and educational purposes only. Not financial advice.