CompactTS-Agent: Autonomous Agentic Meta-Forecaster for Time Series

Developed by Daksh Bhalala.

CompactTS-Agent is an autonomous agentic meta-forecasting system designed for general time-series forecasting across diverse industrial domains (energy grids, server telemetry, traffic flow, nature observations, and healthcare).

Architecture Overview

Rather than relying on a single static neural network across disparate domains, CompactTS-Agent operates as an autonomous orchestration system:

1. Series Profiler: Extracts structural statistical and spectral characteristics from incoming series (Fourier spectral density, dominant seasonal lag, trend variance ratio, and kurtosis).
2. Meta-Routing Policy: Maps the time series to domain-specialized open-weights foundation model representations (TimesFM-3, Chronos-2, Toto-2.0).
3. Calibrated Quantile Fusion: Enforces strict quantile monotonicity ($q_{0.1} \le q_{0.2} \le \dots \le q_{0.9}$) and harmonic basis alignment.

All underlying foundation models utilized in this agentic framework are public open-source weights licensed under Apache-2.0.

Installation & Quickstart

git clone https://github.com/DakshBhalala/AMBACAST.git
cd AMBACAST
pip install -e .

Python API Usage

import numpy as np
import pandas as pd
from compactts.agent import CompactTSAgent, SeriesProfiler

# 1. Profile an incoming time series
profiler = SeriesProfiler(freq="H")
series = np.sin(np.linspace(0, 20, 100)) + 0.1 * np.random.randn(100)
profile = profiler.profile(series)
print("Series Profile:", profile)

# 2. Generate calibrated probabilistic forecasts
agent = CompactTSAgent(prediction_length=24, freq="H")
start_dates = [pd.Period("2024-01-01 00:00", freq="h")]
forecasts = agent.forecast_batch([series], start_dates)

# Access forecast quantiles (9 levels: 0.1 to 0.9)
q_forecasts = forecasts[0].forecast_array
print("Forecast shape (9 quantiles, 24 horizon):", q_forecasts.shape)

Benchmark Evaluation

CompactTS-Agent is evaluated against the 97 dataset configurations of the Salesforce GIFT-Eval benchmark under model_type: "agentic".

To reproduce results:

python scripts/run_agentic_benchmark.py

Citation & Authorship

• Author: Daksh Bhalala
• Code Repository: DakshBhalala/AMBACAST
• License: Apache-2.0