time-series-forecast-transformer

Time-Series-Forecast-Transformer

Time-Series-Forecast-Transformer working at the local container.

Run

docker compose build
docker compose up

Open the frontend at http://127.0.0.1:5173.

The API is also available at http://127.0.0.1:8000.

Backend CLI

Run an offline smoke forecast with the top-level sample dataset:

docker compose run --rm backend \
  python -m local_ts_forecast.cli forecast \
    --backend seasonal_naive \
    --input data/sample_data.csv \
    --target ITEM_A \
    --output outputs/forecast.csv \
    --plot outputs/forecast.png

Run the API example:

docker compose up -d backend
docker compose exec backend python scripts/api_example.py

GPU

docker compose -f docker-compose.yml -f docker-compose.gpu.yml up --build backend

Test

docker compose -f docker-compose.test.yml up --build --abort-on-container-exit

Train

# M4 dataset

# Create long fomart data
docker compose run --rm backend \
  python -m local_ts_forecast.cli convert-dataset \
    --dataset m4 \
    --m4-train-dir data/m4/Train \
    --m4-frequency Monthly \
    --max-series 10000 \
    --output outputs/m4_monthly_train_long.csv

# Create Holdouts
docker compose run --rm backend \
  python scripts/create_holdouts.py


# Chronos-2 zero-shot
docker compose -f docker-compose.yml -f docker-compose.gpu.yml run --rm backend \
  python -m local_ts_forecast.cli forecast \
    --backend chronos2 \
    --input outputs/m4_monthly_train_long.csv \
    --target target \
    --prediction-length 18 \
    --context-length 36 \
    --batch-size 64 \
    --device cuda \
    --output outputs/m4_monthly_zero_shot_chronos2.csv \
    --plot ""

# Chronos-2 adapter/head
docker compose -f docker-compose.yml -f docker-compose.gpu.yml run --rm backend \
  python -m local_ts_forecast.cli train-adapter \
    --dataset m4 \
    --m4-train-dir data/m4/Train \
    --m4-frequency Monthly \
    --base-backend chronos2 \
    --model-id autogluon/chronos-2-small \
    --context-length 36 \
    --prediction-length 18 \
    --max-series 10000 \
    --max-windows 20000 \
    --windows-per-series 2 \
    --base-batch-size 64 \
    --steps 5000 \
    --batch-size 256 \
    --learning-rate 0.001 \
    --hidden-size 128 \
    --device cuda \
    --output outputs/m4_monthly_chronos2_adapter.pt

# adapter
docker compose -f docker-compose.yml -f docker-compose.gpu.yml run --rm backend \
  python -m local_ts_forecast.cli forecast \
    --backend foundation_adapter \
    --checkpoint outputs/m4_monthly_chronos2_adapter.pt \
    --input outputs/m4_monthly_train_long.csv \
    --target target \
    --prediction-length 18 \
    --device cuda \
    --output outputs/m4_monthly_chronos2_adapter_forecast.csv \
    --plot ""

# Compare zero-shot vs adapter
docker compose run --rm backend \
  python scripts/compare_forecast_metrics.py \
    outputs/m4_monthly_zero_shot_chronos2.csv \
    outputs/m4_monthly_chronos2_adapter_forecast.csv

# TimesFM
# zero-shot TimesFM：
docker compose -f docker-compose.yml -f docker-compose.gpu.yml run --rm backend \
  python -m local_ts_forecast.cli forecast \
    --backend timesfm \
    --input outputs/m4_monthly_train_long.csv \
    --target target \
    --prediction-length 18 \
    --context-length 36 \
    --device cuda \
    --output outputs/m4_monthly_zero_shot_timesfm.csv \
    --plot ""

# TimesFM frozen + adapter：
docker compose -f docker-compose.yml -f docker-compose.gpu.yml run --rm backend \
  python -m local_ts_forecast.cli train-adapter \
    --dataset m4 \
    --m4-train-dir data/m4/Train \
    --m4-frequency Monthly \
    --base-backend timesfm \
    --model-id google/timesfm-2.5-200m-pytorch \
    --context-length 36 \
    --prediction-length 18 \
    --max-series 5000 \
    --max-windows 10000 \
    --windows-per-series 2 \
    --base-batch-size 32 \
    --steps 3000 \
    --batch-size 256 \
    --learning-rate 0.001 \
    --hidden-size 128 \
    --device cuda \
    --output outputs/m4_monthly_timesfm_adapter.pt

# TimesFM adapter：
docker compose -f docker-compose.yml -f docker-compose.gpu.yml run --rm backend \
  python -m local_ts_forecast.cli forecast \
    --backend foundation_adapter \
    --checkpoint outputs/m4_monthly_timesfm_adapter.pt \
    --input outputs/m4_monthly_train_long.csv \
    --target target \
    --prediction-length 18 \
    --device cuda \
    --output outputs/m4_monthly_timesfm_adapter_forecast.csv \
    --plot ""

# TimesFM Compare：
docker compose run --rm backend \
  python scripts/compare_forecast_metrics.py \
    outputs/m4_monthly_zero_shot_timesfm.csv \
    outputs/m4_monthly_timesfm_adapter_forecast.csv


# Chronos-2 zero-shot、Chronos-2 adapter、TimesFM zero-shot、TimesFM adapter
docker compose run --rm backend \
  python scripts/compare_forecast_metrics.py \
    outputs/m4_monthly_zero_shot_chronos2.csv \
    outputs/m4_monthly_chronos2_adapter_forecast.csv \
    outputs/m4_monthly_zero_shot_timesfm.csv \
    outputs/m4_monthly_timesfm_adapter_forecast.csv

M4 Monthly benchmark results

Evaluation target: M4 Monthly
Forecast horizon: 18
Evaluation rows: 180,000

Model	Forecast type	Output path	MAE	RMSE	MAPE	sMAPE
Chronos-2	Zero-shot	outputs/m4_monthly_zero_shot_chronos2.csv	598.023868	1374.962430	16.871542%	14.438171%
Chronos-2 + Adapter	Frozen foundation model + trained adapter/head	outputs/m4_monthly_chronos2_adapter_forecast.csv	646.648479	1422.491100	17.943740%	14.863127%
TimesFM	Zero-shot	outputs/m4_monthly_zero_shot_timesfm.csv	607.634425	1367.991547	17.942648%	14.922394%
TimesFM + Adapter	Frozen foundation model + trained adapter/head	outputs/m4_monthly_timesfm_adapter_forecast.csv	765.064199	1540.075852	23.020814%	17.705266%

This site is open source. Improve this page.