Time-Series Forecasting

1. pip install autogluon (AutoGluon — automates time-series forecasting with AutoML, no deep expertise needed).
2. Prepare your data as a CSV with timestamp and target columns. AutoGluon handles feature engineering, model selection, and ensembling.

from autogluon.timeseries import TimeSeriesDataFrame, TimeSeriesPredictor
data = TimeSeriesDataFrame.from_path("sales.csv")
predictor = TimeSeriesPredictor(prediction_length=30, target="revenue").fit(data)
predictions = predictor.predict(data)
print(predictions)

1. First forecast in 5-15 minutes on CPU for a moderate dataset (10K rows, 50 time series). AutoGluon trains an ensemble of statistical (ARIMA, ETS) + ML (LightGBM, CatBoost) + deep learning (DeepAR, PatchTST) models.
2. For foundation model-based forecasting: pip install chronos (Amazon Chronos — time-series foundation model). Zero-shot forecasting: feed any time series → Chronos predicts without training on your data. First forecast in seconds on GPU.
3. For deep learning forecasting: pip install darts (u8darts, unit8) — supports TFT, N-BEATS, N-HiTS, TiDE architectures with a clean API.
4. For LLM-based forecasting: pip install lag-llama (Lag-Llama) or TimeGPT (proprietary). LLM-based forecasters treat time series as tokens.

Time-series forecasting is CPU-friendly. AutoGluon trains on 10K rows in 5-15 minutes on a $300 laptop. Chronos-tiny (100M params) runs on CPU at 100-500 data points/second — forecast 1,000 time series in seconds. For enterprise-scale forecasting (100K+ SKUs, daily forecasts): Chronos-small (350M params) on a used GTX 1060 6 GB ($60) handles it in minutes. Total: ~$360. Forecasting at $300-400 handles 90% of business use cases (retail demand, inventory, capacity planning). The bottleneck is data quality and feature engineering, not compute.

Used RTX 3060 12 GB ($200-250, see /hardware/rtx-3060-12gb) handles Chronos-large (710M params) for high-accuracy zero-shot forecasting on 100K+ time series. For training custom deep forecasting models (PatchTST, TimesNet) on large-scale retail data (1M+ SKUs): GPU training in 4-12 hours. For production forecasting pipelines with automated retraining, ensemble selection, and backtesting: the GPU accelerates the training phase; inference runs on CPU. Total: ~$800-1,000. Forecasting compute needs are modest — invest in data infrastructure (pipeline, quality monitoring) before GPU.

The mistake: Training a sophisticated deep learning forecaster (Transformer, N-BEATS) on a time series with 200 data points, then comparing it to a 3-line exponential smoothing baseline — and the baseline wins. Why it fails: Deep learning models need data — a Transformer with 1M parameters can't learn meaningful patterns from 200 data points. It overfits to noise. Meanwhile, Holt-Winters exponential smoothing has 3 parameters and 60 years of statistical theory behind it — it's near-optimal for short, trend-seasonal series. The fix: Always run simple baselines first: (1) naive (last value repeated), (2) seasonal naive (same period last year), (3) exponential smoothing, (4) ARIMA. Only if a deep model beats all four baselines on a held-out test set should you consider deploying it. For short series (<500 points), statistical methods almost always win. For long series (10K+ points) with complex patterns, deep learning pulls ahead. Match model complexity to data quantity.