forecasting-v3-checkpoints

Forecasting V3 Checkpoints

This repository preserves the three final PyTorch checkpoints from the forecasting-v3 AAPL/ABBV/MCD development study.

Model

• One shared model for AAPL, ABBV, and MCD
• 128 completed one-minute close tokens
• Continuous RoPE horizon queries with base 16
• Supervised horizons: 2, 4, 8, 16, and 32 minutes
• Dense query support: every integer horizon from 2 through 32 minutes
• Shared 21-class return-distribution output
• Separate 15-feature causal long-history auxiliary lane
• 27,477 trainable parameters

Each seed was initialized from its locked seed-matched early-add checkpoint, trained end to end on [2016-08-08, 2021-08-08), selected with [2021-08-08, 2022-08-08), and refit on [2016-08-08, 2022-08-08). The three selected refit epoch counts were 7, 1, and 4.

The strategy-development interval was [2022-08-08, 2023-08-08). The sealed reserve [2023-08-08, 2026-08-08) was not read or scored.

Files

Each seedN/ directory contains:

• model_seedN.pt: weights and the core architecture metadata
• calibration_seedN.npz: class-return supports and train/refit calibration
• results.json: training history and development metrics
• run_manifest.json: exact configuration and source hashes

source/ contains the model definitions and frozen V7 runner used to create the checkpoints. protocol.md records the study contract. Prediction sheets and market data are intentionally excluded.

Load

import torch

checkpoint = torch.load(
    "seed0/model_seed0.pt",
    map_location="cpu",
    weights_only=True,
)
state_dict = checkpoint["state_dict"]
print(checkpoint["assets"])
print(checkpoint["supervised_horizons"])

Use the source and manifest files to reconstruct the module before loading the state dictionary.

Development Metrics

Dense 2-32 minute development-test NLL:

These are research artifacts, not investment advice or a production trading system. The development year had prior generic inspection and is not an untouched confirmatory holdout.