Stock Prediction, Chronos, Backtesting, and Market Data

Configure stock prediction with Chronos models, backtesting, and market data sources

Stock prediction provides daily, hourly, and 15‑minute forecasts (Chronos or technical strategies), backtesting over OHLCV, order recommendations, and market-status. It depends on market data (Alpaca or yahooquery) and optionally Chronos via Modal or locally. Rolling credits (stock_prediction_daily, stock_prediction_hourly, stock_prediction_15min) apply when RollingCreditsService is used.

---

1. Overview

Component	Role
StockPredictionService	predict_daily, predict_hourly, predict_15min — Chronos or technical; uses ChronosModelManager, MarketDataService, credits
ChronosModelManager	Chronos inference: Modal chronos_inference or local chronos-bolt when STOCK_PREDICTION_USE_LOCAL=true
MarketDataService	OHLCV: Alpaca (if ALPACA_DATA_ENABLED) or yahooquery (fallback). Used by prediction and backtesting
Backtesting	run_backtest_from_data_source in app.stock_prediction_core.backtesting; strategies: trend, mean_reversion, momentum, volatility, stat_arb, combined
Modal	chronos_inference, chronos_train (stub), market_status — optional when not using local Chronos

Code: app/services/stock_prediction_service.py, app/services/chronos_model_manager.py, app/services/market_data_service.py, app/stock_prediction_core/backtesting.py, modal_app/app.py, app/api/stock_prediction_routes.py

---

2. Configuration

2.1 Feature and Lookback

Variable	Description	Default
STOCK_PREDICTION_ENABLED	Enable prediction APIs (daily, hourly, 15min, backtest, recommend-order, models, market-status). If false, all return 403.	false
STOCK_PREDICTION_DEFAULT_LOOKBACK_DAILY	Default lookback bars for daily	252
STOCK_PREDICTION_DEFAULT_LOOKBACK_HOURLY	Default lookback bars for hourly	504
STOCK_PREDICTION_DEFAULT_LOOKBACK_15MIN	Default lookback bars for 15min	96

2.2 Market Data (OHLCV for Prediction and Backtesting)

Market data is provided by MarketDataService: Alpaca first (when enabled), then yahooquery.

Source	When used	Requirements
Alpaca	ALPACA_DATA_ENABLED=true and ALPACA_API_KEY / ALPACA_API_SECRET set	alpaca-py installed; keys from Alpaca. See Alpaca Trading Setup.
yahooquery	Fallback when Alpaca is off or returns no data	pip install yahooquery

Config (Alpaca as data source):

Variable	Description	Default
ALPACA_DATA_ENABLED	Use Alpaca for historical bars in stock prediction and backtesting	false
ALPACA_API_KEY	Alpaca API key	—
ALPACA_API_SECRET	Alpaca API secret	—

Same keys are used for trading (place/cancel orders, portfolio). Base URL: ALPACA_BASE_URL (e.g. https://paper-api.alpaca.markets for paper).

2.3 Chronos: Modal vs Local

Variable	Description	Default
STOCK_PREDICTION_USE_LOCAL	true: run Chronos locally with chronos-bolt; false: use Modal chronos_inference	false
CHRONOS_MODEL_ID	Model id (e.g. amazon/chronos-t5-small, amazon/chronos-t5-base). Selectable in Predictions tab and via model_id query.	amazon/chronos-t5-small
CHRONOS_DEVICE	Device for Chronos: cpu, cuda, cuda:0. Used by Modal and when running locally.	cpu

Local Chronos (STOCK_PREDICTION_USE_LOCAL=true):

pip install chronos-bolt torch

Modal (STOCK_PREDICTION_USE_LOCAL=false):

Variable	Description	Default
MODAL_APP_NAME	Modal app name	creditnexus-stock-prediction
MODAL_TOKEN_ID	Modal token ID for server-side client (optional)	—
MODAL_TOKEN_SECRET	Modal token secret	—
MODAL_USE_GPU	Env at run/deploy time. 1, true, or yes: Modal chronos_inference and chronos_train use GPU (T4). Set when running: MODAL_USE_GPU=1 modal run modal_app/app.py or MODAL_USE_GPU=1 modal deploy	—

Modal image: modal_app/image.py (chronos-bolt, torch, alpaca-py, pandas, etc.).

---

3. APIs and UI

Endpoint	Method	Description
/api/stock-prediction/daily	GET	Daily prediction; symbol, lookback, horizon, strategy, model_id
/api/stock-prediction/hourly	GET	Hourly prediction
/api/stock-prediction/15min	GET	15‑minute prediction
/api/stock-prediction/backtest	POST	Backtest; body: symbol, start, end, strategy, timeframe, initial_capital; returns equity_curve, trades
/api/stock-prediction/models	GET	Selectable Chronos model ids
/api/stock-prediction/market-status	GET	Market status (market=US_STOCKS, etc.)
/api/stock-prediction/recommend-order	POST	Order recommendation; requires auth

UI: Trading Dashboard → Predictions (run prediction, Chronos model dropdown, chart, order recommendation), Backtest (form, equity curve, trades table).

---

4. Backtesting

• Strategies: combined, trend, mean_reversion, momentum, volatility, stat_arb
• Timeframes: 1d, 1h, 15m (mapped to 1D, 1H, 15Min for data)
• Data: MarketDataService.get_historical_data (Alpaca or yahooquery). Requires ≥130 bars; for 1d, typically ≥6 months of history.
• Request: symbol, start, end (YYYY‑MM‑DD), strategy, timeframe, initial_capital (default 100_000)
• Response: total_return, sharpe_ratio, max_drawdown, win_rate, n_trades, equity_curve, trades, metadata

---

5. Rolling Credits (Stock Prediction)

When RollingCreditsService is used, stock_prediction_daily, stock_prediction_hourly, stock_prediction_15min are deducted. Tier allocation is in Rolling Credits Setup. If the user has no balance, prediction may be refused depending on service wiring.

---

6. Quick Start

1. Set STOCK_PREDICTION_ENABLED=true.
2. Market data: either
   • ALPACA_DATA_ENABLED=true and ALPACA_API_KEY, ALPACA_API_SECRET (and ALPACA_BASE_URL for trading), or
   • Rely on yahooquery (pip install yahooquery).
3. Chronos:
   • Modal: pip install modal, modal token new; deploy/run with MODAL_USE_GPU=1 if you want GPU. Set MODAL_TOKEN_ID/MODAL_TOKEN_SECRET only if the server needs to spawn Modal runs.
   • Local: STOCK_PREDICTION_USE_LOCAL=true, pip install chronos-bolt torch, set CHRONOS_DEVICE (e.g. cuda if GPU).
4. Open Trading → Predictions and Backtest.

---

7. Troubleshooting

Issue	Check
403 on prediction/backtest	STOCK_PREDICTION_ENABLED=true
”Insufficient data” in backtest	Date range and symbol; need ≥130 bars. For 1d, use ~6+ months. yahooquery/Alpaca must return OHLCV.
Chronos “chronos/torch not installed”	Local: pip install chronos-bolt torch
Modal inference fails	modal run modal_app/app.py or modal deploy; MODAL_TOKEN_ID/MODAL_TOKEN_SECRET if the app invokes Modal. MODAL_USE_GPU=1 when deploying for GPU.
No market data	yahooquery installed; or Alpaca with ALPACA_DATA_ENABLED, valid keys, and ALPACA_BASE_URL when also trading.