wti-lstm-autoencoder

WTI LSTM autoencoder

Unsupervised reconstruction-error model for daily WTI futures (CL=F). The network reconstructs 10-day windows of causally vol-normalized ΔClose (USD/bbl). A window is flagged when its reconstruction MSE exceeds the 99th percentile of quiet calibration MSE (2010–2019). That threshold is frozen at 0.382. 2008 is a historical holdout (not in the loss). 2020–present is out of sample.

This is a shape-break score, not a crisis classifier and not a forecast. A rolling 10-day vol rule and a one-day robust z-score are published next to it at the same P99 budget. Training and scoring live in DrAdrianDC/WTI_Anomaly_Detection.

Files

Contract

• Ticker: CL=F, unadjusted close
• Input: (batch, 10, 1) of ( x_t = \Delta\mathrm{Close}_t / \max(\mathrm{causal\ 60d\ MAD},\ 0.66) )
• Percent/log returns are not used (negative print, 20 April 2020)
• Threshold: P99 of quiet 2010–2019 train MSE = 0.382, frozen
• Weights are not updated on a schedule

Champion snapshot

• last_date: 2026-09-10
• Calibration: 1,974 quiet train + 349 early-stop; 121 loud windows dropped
• Train MAE 0.187; early-stop MAE 0.287; best epoch 194 / 200
• Flags: 368 / 6,470 (5.7%); OOS 2020–2026: 174 / 1,683 (10.3%)
• OOS catalog (10 pre-registered events): LSTM 9 / 10, rolling vol 8 / 10
• Spearman(MSE, 10-day vol): 0.33
• 20 April 2020: close −37.63 USD, (M=2.44); peak ringing 28 April (M=2.78)
• June 2022 liquidation: miss ((M=-1.23)). Vol baseline hits. Grind after local MAD has risen
• May–September 2026: 0 LSTM flags after the March–April episode

P99 is of quiet train MSE, not of live days. Rank (M=\log_{10}(\mathrm{MSE}/\mathrm{threshold})); do not treat the 0/1 as a 1% live rate.

Use

import json
from huggingface_hub import hf_hub_download
from tensorflow import keras

repo_id = "DrAdrianDC/wti-lstm-autoencoder"
model = keras.models.load_model(hf_hub_download(repo_id, "lstm_autoencoder.keras"))
state = json.loads(open(hf_hub_download(repo_id, "feature_state.json")).read())
threshold = json.load(open(hf_hub_download(repo_id, "metadata.json")))["threshold"]

# Build windows with the GitHub package (`src.features.build_feature_frame`)
# so the causal MAD matches training. Do not use a global scaler.

There is no Transformers pipeline() for this checkpoint.

Limitations

• Not a forecast. The window has already ended.
• Not a crisis classifier. 2014–16 glut (in calibration) and June 2022 do not flag.
• Univariate unadjusted close only.
• Changing lookback, vol lookback, or activation invalidates the champion; run train.

Intended use

Research and portfolio demonstration. Not for automated trading. If the flag fires on a live print, check CME before the model.