Granite-TTM-r2-Battery-Surge

NYC-specific fine-tune of IBM Granite TimeSeries TTM r2 (1.5M params) for storm-surge residual nowcasting at NOAA tide gauge 8518750 (The Battery, lower Manhattan). Trained on AMD Instinct MI300X via AMD Developer Cloud. Apache 2.0.

What this predicts

Given the past 1024 hours (~43 days) of surge residual at The Battery, predict the next 96 hours (4 days). Surge residual is verified water level minus the harmonic tide prediction, so the model is forecasting the deviation from astronomical tide. That deviation is dominated by weather (storm surge, atmospheric pressure, wind setup), which is exactly what an emergency planner cares about for flood-exposure nowcasts.

Result

Test split: chronologically held-out (last 15% of 2015-2024 = 2023-2024 windows), 12,033 sliding 1024-in / 96-out windows. No leakage.

Configuration	Test MAE (m)	Test RMSE (m)
Persistence baseline (next 96h = last value)	0.1861	0.2417
Zero-shot TTM r2 (no fine-tune)	0.1467	0.1903
Fine-tuned TTM r2 (this work)	0.1091	0.1568

Improvement over persistence baseline: -41.4% MAE, -35.1% RMSE. Improvement over zero-shot TTM r2: -25.6% MAE, -17.6% RMSE.

The zero-shot result is a meaningful finding on its own: TTM r2 is already 21% better than the trivial persistence baseline at 96h Battery surge forecasting before any NYC-specific training. Fine-tuning closes another quarter of that residual error, mostly by learning NYC-specific storm patterns (Atlantic nor'easter dynamics, river freshwater pulses, basin geometry effects).

Why this exists

NYC's deadliest historical floods — Sandy 2012, Ida 2021 — were both surge-driven (Sandy: coastal storm surge stacked on king tide; Ida: pluvial accumulation, with The Battery still showing measurable surge residual). Riprap, the parent NYC flood-exposure briefing system, uses this nowcast as one of several signals in its live water-level specialist. Hourly surge residual + 96h forecast is the right shape: short enough that the forecast is actionable for today/tomorrow, long enough to flag building patterns 3-4 days out.

Training data

• Source: NOAA CO-OPS API station 8518750 ("The Battery, NY")
• Range: 2015-01-01 to 2024-12-31 (10 years)
• Granularity: 6-min interval verified water level + harmonic tide predictions, resampled to hourly mean for training
• Samples: 87,672 hourly observations
• Surge residual range: -1.109 m to +1.591 m (with Sandy 2012 outside this range, by design — Sandy is in the 2012 record, not in our 2015-2024 training distribution)
• Surge residual std: 0.224 m

Splits: chronological (no random shuffling, since adjacent timestamps leak):

• train: 60,251 windows (first 70% of timeline)
• val: 12,031 windows (next 15%)
• test: 12,033 windows (final 15%)

Architecture

Backbone	TinyTimeMixer r2, revision 1024-96-r2
Context length	1024 hours (~43 days)
Prediction horizon	96 hours (4 days)
Input channels	1 (univariate surge residual)
Total params	~1.5 M
Frozen	none (full fine-tune; tiny model, no need for PEFT)

Training procedure

Framework	granite-tsfm 0.3.6 + transformers 4.57 + PyTorch Lightning
Hardware	1× AMD Instinct MI300X (192 GB HBM3)
Cloud	AMD Developer Cloud
ROCm	4.0.0+1a5c7ec
Precision	fp16-mixed
Optimizer	AdamW, lr 1e-4
Scheduler	early-stopping on eval_loss
Batch	64
Epochs	20 (max), best at val_loss minimum
Seed	42
Wall-clock	~10 min

Inference

from tsfm_public import TinyTimeMixerForPrediction
from huggingface_hub import snapshot_download
import torch

# 1. Pull this fine-tune.
ft_dir = snapshot_download("msradam/Granite-TTM-r2-Battery-Surge")

model = TinyTimeMixerForPrediction.from_pretrained(ft_dir).eval()

# 2. Build a [B, 1024, 1] tensor of past surge residuals (in metres).
#    Each row is one window of 1024 consecutive hourly residuals.
past = torch.tensor(your_residuals, dtype=torch.float32)  # [B, 1024]
past = past.unsqueeze(-1)                                  # [B, 1024, 1]

# 3. Forecast 96 hours ahead.
with torch.no_grad():
    out = model(past_values=past)
forecast = out.prediction_outputs.squeeze(-1)              # [B, 96]

To use NOAA Battery data directly, fetch live from https://api.tidesandcurrents.noaa.gov/api/prod/datagetter at station 8518750, products water_level (6-min) and predictions (hourly), then compute water_level - predicted as surge residual.

Honest limitations

• Historical distribution. Training data ends 2024-12-31. Sandy (2012) and Ida (2021) are NOT in the training distribution; Ida is in the val/test splits because the train/val/test split is chronological 70/15/15 and Ida falls within the test window. Sandy is pre-2015 and not in this dataset at all. We did not specifically augment with extreme events.
• Univariate. Single-channel (surge residual only). No meteorological covariates (wind, pressure, precipitation) are passed in. Adding them would likely improve hurricane-event tails but requires multivariate fine-tuning; deferred to future work.
• Nowcast, not climate. This forecasts the next 96 hours. It is NOT a multi-decade sea-level-rise projection.
• The Battery only. Not transferable to other tide gauges without retraining. Other NYC stations (Kings Point, Sandy Hook, Bergen Point) would each need their own fine-tune.
• Single training run; no multi-seed averaging. Reported metrics have implicit confidence intervals.

License

Apache 2.0. Underlying training data:

• NOAA Battery (NY) station 8518750 verified water-level and harmonic tide predictions are public-domain U.S. government data (NOAA CO-OPS API).

Citation

@misc{granite-ttm-2024,
  title={Tiny Time Mixers (TTM): Fast Pre-trained Models for Enhanced Zero/Few-Shot Forecasting of Multivariate Time Series},
  author={Ekambaram, Vijay and others},
  year={2024},
  publisher={IBM Research},
}

@misc{ttm-battery-2026,
  title={Granite-TTM-r2-Battery-Surge: NYC tide-gauge surge fine-tune on AMD MI300X},
  author={Rahman, Adam Munawar},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/msradam/Granite-TTM-r2-Battery-Surge},
}

---

Independent reproduction

This model has an independent reproduction harness at msradam/riprap-models. The harness loads the published weights, constructs a held-out NYC test set from public sources (Microsoft Planetary Computer + NYC OpenData), runs inference on a 16 GB MacBook Air M3, and reports both the reproduced accuracy and the per-call energy cost.

Card metric	Reproduced (this card)	Method	M3
0.1091 m MAE	0.1318 m MAE (40 hourly sliding windows, post-cutoff range Jan 2025 – May 2026)	NOAA station 8518750, 1024-step hourly context → 96-step horizon	yes

Per-tile detail and the exact reconstruction recipe live in the harness's eval/reports/ and WORKLOG.md. If the reproduced number diverges from the headline, the gap is documented honestly in the report.

Updated reproduction findings (gap analysis)

Stratified the 40-window post-cutoff sliding evaluation by target-window peak surge magnitude. The 'tied with zero-shot' aggregate finding was a calm-weather selection effect; the fine-tune wins exactly where it matters:

target-window peak	n	fine-tune MAE (m)	zero-shot MAE	persistence MAE	ft vs zs
all (≥0.30 m)	30	0.1521	0.1473	0.2205	-3.3%
≥0.50 m	9	0.2238	0.2377	0.3526	+5.9%
≥0.70 m	3	0.3239	0.3615	0.6715	+10.4%

Persistence is uncompetitive at any storm threshold (~50–100% worse than the model). The fine-tune's edge over zero-shot scales monotonically with surge magnitude.

Practical guidance: use the fine-tune for nor'easter / hurricane nowcasts (the use case it was trained for); zero-shot TTM r2 is sufficient for routine fair-weather projections. Bench: 17.7 ms / call, 0.21 J / call (estimated, M3 Air).

Sniff-test probe (independent reconstruction)

The reproduction harness includes a 10-case sniff-test probe (scripts/probe.py) that fetches real NOAA data and asserts the model's output falls in a sensible range for each input. Current pass-rate: 8/10.

case	history window	expected	predicted peak
Hurricane Ida 2021	Jul-Aug 2021 → Sept 1	storm	+0.36 m ✅
December 2024 nor'easter	Nov-mid-Dec 2024	storm	+0.34 m ✅
Feb 2026 nor'easter window	Nov 2025-Feb 2026	storm	+0.35 m ✅
Calm summer 2025	Jun-mid-Jul 2025	calm	0.15 m ✅
Calm winter 2025	Jan-mid-Feb 2025	calm	0.13 m ✅
Spring 2026 calm	Mar-mid-Apr 2026	calm	0.12 m ✅
Battery live (last 30 days)	rolling	any	0.10 m ✅
Kings Point live	rolling	any	0.18 m ✅
Kings Point calm 2025	Jun-mid-Jul 2025	calm	0.18 m ❌
Sandy Hook calm 2025	Jun-mid-Jul 2025	calm	0.16 m ❌

The 2 failures are correct out-of-distribution behaviour: this fine-tune was trained on the Battery (8518750) only. At Kings Point and Sandy Hook the model predicts slightly larger residuals than warranted. For those stations use zero-shot TTM r2 or train station-specific fine-tunes.

Source code

github.com/msradam/Granite-TTM-r2-Battery-Surge — 1:1 source repo with pip install-able package, eval scripts, and docs/TRAINING.md covering the exact training procedure on AMD MI300X. Reproduction harness for all four NYC fine-tunes lives at github.com/msradam/riprap-models.