Skyloom Karachi AQI Forecasting

This is a fine‑tuned version of google/timesfm-2.5-200m-transformers for hourly Air Quality Index (AQI) forecasting in Karachi, Pakistan.
The model is optimised for a 24‑step forecast horizon (next 24 hours).

Model summary

Item	Value
Base model	google/timesfm-2.5-200m-transformers
Task	Univariate time‑series forecasting
Target variable	Hourly AQI
Location	Karachi, Pakistan
Forecast horizon (training)	24 hours
Forecast horizon (evaluation)	24 hours
Test MAE	5.178
Test sMAPE	5.456%
Test MASE	0.4687

Data

• Source: Syed110-3/karachi-aqi-predictor
• After cleaning: 26,208 hourly records (2023-01-02 to 2025-12-28)
• Split: chronological (no future leakage)
• Seasonal period for MASE: 24 hours

Evaluation results (24‑step ahead)

Model	MAE	RMSE	sMAPE	MASE
Skyloom	5.178	8.859	5.456	0.4687
ValidationWeightedEnsemble	5.170	8.824	5.449	0.4679
TimesFM‑ZeroShot	5.279	9.023	5.546	0.4778
HGBDirectLagCalendar	6.899	10.597	7.390	0.6244
SeasonalDrift24h	10.238	15.391	10.867	0.9266
RollingMean24h	10.749	15.926	11.281	0.9729
SeasonalNaive	11.008	16.520	11.629	0.9962
SameHourMean7D	13.962	19.036	15.142	1.2636
ETS_Additive24h	13.962	19.036	15.142	1.2636

Note: The ensemble assigned 90% weight to the model and 0% to the zero‑shot model, confirming that fine‑tuning was essential for this dataset.

Usage

import torch
import numpy as np
from transformers import TimesFm2_5ModelForPrediction

model_id = "mahwizzzz/skyloom"  
device = "cuda" if torch.cuda.is_available() else "cpu"

model = TimesFm2_5ModelForPrediction.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16 if device == "cuda" else torch.float32,
    device_map=device
)
model.eval()

# Provide the last 1024 hourly AQI values
past_aqi = np.array([...], dtype=np.float32)

with torch.no_grad():
    batch = torch.tensor(past_aqi[-1024:][None, :], dtype=torch.float32).to(device)
    output = model(past_values=batch, forecast_context_len=1024)
    forecast_24h = output.mean_predictions[0, :24].float().cpu().numpy()

forecast_24h = np.clip(forecast_24h, 0, None)
print(forecast_24h)

Limitations

• Only AQI history is used no weather, traffic, or other covariates.
• Performance may degrade during unexpected pollution events or sensor anomalies.
• The reported metrics are from historical backtesting, not live operation.

Citation

If you use this model, cite the base TimesFM work and this model repository:

@misc{skyloom2026,
  title        = {Skyloom: TimesFM 2.5 for Karachi AQI Forecasting},
  author       = {Mahwiz khalil},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/mahwizzzz/skyloom}}
}