Zero-shot time series forecasting with Google's TimesFM foundation model. Use for any univariate time series (sales, sensors, energy, vitals, weather) without training a custom model. Supports CSV/DataFrame/array inputs with point forecasts and prediction intervals. Includes a preflight system checker script to verify RAM/GPU before first use.
Zero-shot time series forecasting with Google's TimesFM foundation model. Use for any univariate time series (sales, sensors, energy, vitals, weather) without training a custom model. Supports CSV/DataFrame/array inputs with point forecasts and prediction intervals. Includes a preflight system checker script to verify RAM/GPU before first use.
allowed-tools
Read Write Edit Bash
license
Apache-2.0 license
metadata
{"skill-author":"Clayton Young / Superior Byte Works, LLC (@borealBytes)","skill-version":"1.0.0"}
TimesFM Forecasting
Overview
TimesFM (Time Series Foundation Model) is a pretrained decoder-only foundation model
developed by Google Research for time-series forecasting. It works zero-shot — feed it
any univariate time series and it returns point forecasts with calibrated quantile
prediction intervals, no training required.
This skill wraps TimesFM for safe, agent-friendly local inference. It includes a
mandatory preflight system checker that verifies RAM, GPU memory, and disk space
before the model is ever loaded so the agent never crashes a user's machine.
Key numbers: TimesFM 2.5 uses 200M parameters (~800 MB on disk, ~1.5 GB in RAM on
CPU, ~1 GB VRAM on GPU). The archived v1/v2 500M-parameter model needs ~32 GB RAM.
Always run the system checker first.
When to Use This Skill
Use this skill when:
Forecasting any univariate time series (sales, demand, sensor, vitals, price, weather)
You need zero-shot forecasting without training a custom model
You want probabilistic forecasts with calibrated prediction intervals (quantiles)
You have time series of any length (the model handles 1–16,384 context points)
You need to batch-forecast hundreds or thousands of series efficiently
You want a foundation model approach instead of hand-tuning ARIMA/ETS parameters
Do not use this skill when:
You need classical statistical models with coefficient interpretation → use statsmodels
You need time series classification or clustering → use aeon
You need multivariate vector autoregression or Granger causality → use statsmodels
Your data is tabular (not temporal) → use scikit-learn
Note on Anomaly Detection: TimesFM does not have built-in anomaly detection, but you can
use the quantile forecasts as prediction intervals — values outside the 90% CI (q10–q90)
are statistically unusual. See the examples/anomaly-detection/ directory for a full example.
⚠️ Mandatory Preflight: System Requirements Check
CRITICAL — ALWAYS run the system checker before loading the model for the first time.
python scripts/check_system.py
This script checks:
Available RAM — warns if below 4 GB, blocks if below 2 GB
GPU availability — detects CUDA/MPS devices and VRAM
Disk space — verifies room for the ~800 MB model download
Python version — requires 3.10+
Existing installation — checks if timesfm and torch are installed
Note: Model weights are NOT stored in this repository. TimesFM weights (~800 MB)
download on-demand from HuggingFace on first use and cache in ~/.cache/huggingface/.
The preflight checker ensures sufficient resources before any download begins.
flowchart TD
accTitle: Preflight System Check
accDescr: Decision flowchart showing the system requirement checks that must pass before loading TimesFM.
start["🚀 Run check_system.py"] --> ram{"RAM ≥ 4 GB?"}
ram -->|"Yes"| gpu{"GPU available?"}
ram -->|"No (2-4 GB)"| warn_ram["⚠️ Warning: tight RAM<br/>CPU-only, small batches"]
ram -->|"No (< 2 GB)"| block["🛑 BLOCKED<br/>Insufficient memory"]
warn_ram --> disk
gpu -->|"CUDA / MPS"| vram{"VRAM ≥ 2 GB?"}
gpu -->|"CPU only"| cpu_ok["✅ CPU mode<br/>Slower but works"]
vram -->|"Yes"| gpu_ok["✅ GPU mode<br/>Fast inference"]
vram -->|"No"| cpu_ok
gpu_ok --> disk{"Disk ≥ 2 GB free?"}
cpu_ok --> disk
disk -->|"Yes"| ready["✅ READY<br/>Safe to load model"]
disk -->|"No"| block_disk["🛑 BLOCKED<br/>Need space for weights"]
classDef ok fill:#dcfce7,stroke:#16a34a,stroke-width:2px,color:#14532d
classDef warn fill:#fef9c3,stroke:#ca8a04,stroke-width:2px,color:#713f12
classDef block fill:#fee2e2,stroke:#dc2626,stroke-width:2px,color:#7f1d1d
classDef neutral fill:#f3f4f6,stroke:#6b7280,stroke-width:2px,color:#1f2937
class ready,gpu_ok,cpu_ok ok
class warn_ram warn
class block,block_disk block
class start,ram,gpu,vram,disk neutral
Hardware Requirements by Model Version
Model
Parameters
RAM (CPU)
VRAM (GPU)
Disk
Context
TimesFM 2.5 (recommended)
200M
≥ 4 GB
≥ 2 GB
~800 MB
up to 16,384
TimesFM 2.0 (archived)
500M
≥ 16 GB
≥ 8 GB
~2 GB
up to 2,048
TimesFM 1.0 (archived)
200M
≥ 8 GB
≥ 4 GB
~800 MB
up to 2,048
Recommendation: Always use TimesFM 2.5 unless you have a specific reason to use an
older checkpoint. It is smaller, faster, and supports 8× longer context.
🔧 Installation
Step 1: Verify System (always first)
python scripts/check_system.py
Step 2: Install TimesFM
# Using uv (recommended by this repo)
uv pip install timesfm[torch]
# Or using pip
pip install timesfm[torch]
# For JAX/Flax backend (faster on TPU/GPU)
uv pip install timesfm[flax]
Step 3: Install PyTorch for Your Hardware
# CUDA 12.1 (NVIDIA GPU)
pip install torch>=2.0.0 --index-url https://download.pytorch.org/whl/cu121
# CPU only
pip install torch>=2.0.0 --index-url https://download.pytorch.org/whl/cpu
# Apple Silicon (MPS)
pip install torch>=2.0.0 # MPS support is built-in
flowchart LR
accTitle: Quantile Forecast Anatomy
accDescr: Diagram showing how the 10-element quantile vector maps to prediction intervals.
input["📈 Input Series<br/>1-D array"] --> model["🤖 TimesFM<br/>compile + forecast"]
model --> point["📍 Point Forecast<br/>(batch, horizon)"]
model --> quant["📊 Quantile Forecast<br/>(batch, horizon, 10)"]
quant --> pi80["80% PI<br/>q[:,:,1] – q[:,:,9]"]
quant --> pi60["60% PI<br/>q[:,:,2] – q[:,:,8]"]
quant --> median["Median<br/>q[:,:,5]"]
classDef data fill:#dbeafe,stroke:#2563eb,stroke-width:2px,color:#1e3a5f
classDef model fill:#f3e8ff,stroke:#9333ea,stroke-width:2px,color:#581c87
classDef output fill:#dcfce7,stroke:#16a34a,stroke-width:2px,color:#14532d
class input data
class model model
class point,quant,pi80,pi60,median output
🔧 ForecastConfig Reference
All forecasting behavior is controlled by timesfm.ForecastConfig:
timesfm.ForecastConfig(
max_context=1024, # Max context window (truncates longer series)
max_horizon=256, # Max forecast horizon
normalize_inputs=True, # Normalize inputs (RECOMMENDED for stability)
per_core_batch_size=32, # Batch size per device (tune for memory)
use_continuous_quantile_head=True, # Better quantile accuracy for long horizons
force_flip_invariance=True, # Ensures f(-x) = -f(x) (mathematical consistency)
infer_is_positive=True, # Clamp forecasts ≥ 0 when all inputs > 0
fix_quantile_crossing=True, # Ensure q10 ≤ q20 ≤ ... ≤ q90
return_backcast=False, # Return backcast (for covariate workflows)
)
Parameter
Default
When to Change
max_context
0
Set to match your longest historical window (e.g., 512, 1024, 4096)
max_horizon
0
Set to your maximum forecast length
normalize_inputs
False
Always set True — prevents scale-dependent instability
per_core_batch_size
1
Increase for throughput; decrease if OOM
use_continuous_quantile_head
False
Set True for calibrated prediction intervals
force_flip_invariance
True
Keep True unless profiling shows it hurts
infer_is_positive
True
Set False for series that can be negative (temperature, returns)
fix_quantile_crossing
False
Set True to guarantee monotonic quantiles
📋 Common Workflows
Workflow 1: Single Series Forecast
flowchart TD
accTitle: Single Series Forecast Workflow
accDescr: Step-by-step workflow for forecasting a single time series with system checking.
check["1. Run check_system.py"] --> load["2. Load model<br/>from_pretrained()"]
load --> compile["3. Compile with ForecastConfig"]
compile --> prep["4. Prepare data<br/>pd.read_csv → np.array"]
prep --> forecast["5. model.forecast()<br/>horizon=N"]
forecast --> extract["6. Extract point + PI"]
extract --> plot["7. Plot or export results"]
classDef step fill:#f3f4f6,stroke:#6b7280,stroke-width:2px,color:#1f2937
class check,load,compile,prep,forecast,extract,plot step
import pandas as pd, numpy as np
# Load wide-format CSV (one column per series)
df = pd.read_csv("all_stores.csv", parse_dates=["date"], index_col="date")
inputs = [df[col].dropna().values.astype(np.float32) for col in df.columns]
# Forecast all series at once (batched internally)
point, quantiles = model.forecast(horizon=30, inputs=inputs)
# Collect results
results = {}
for i, col inenumerate(df.columns):
results[col] = {
"forecast": point[i].tolist(),
"lower_80": quantiles[i, :, 1].tolist(),
"upper_80": quantiles[i, :, 9].tolist(),
}
# Exportimport json
withopen("batch_forecasts.json", "w") as f:
json.dump(results, f, indent=2)
print(f"Forecasted {len(results)} series → batch_forecasts.json")
Workflow 3: Evaluate Forecast Accuracy
import numpy as np
# Hold out the last H points for evaluation
H = 24
train = values[:-H]
actual = values[-H:]
point, quantiles = model.forecast(horizon=H, inputs=[train])
pred = point[0]
# Metrics
mae = np.mean(np.abs(actual - pred))
rmse = np.sqrt(np.mean((actual - pred) ** 2))
mape = np.mean(np.abs((actual - pred) / actual)) * 100# Prediction interval coverage
lower = quantiles[0, :, 1]
upper = quantiles[0, :, 9]
coverage = np.mean((actual >= lower) & (actual <= upper)) * 100print(f"MAE: {mae:.2f}")
print(f"RMSE: {rmse:.2f}")
print(f"MAPE: {mape:.1f}%")
print(f"80% PI Coverage: {coverage:.1f}% (target: 80%)")
⚙️ Performance Tuning
GPU Acceleration
import torch
# Check GPU availabilityif torch.cuda.is_available():
print(f"GPU: {torch.cuda.get_device_name(0)}")
print(f"VRAM: {torch.cuda.get_device_properties(0).total_mem / 1e9:.1f} GB")
elifhasattr(torch.backends, "mps") and torch.backends.mps.is_available():
print("Apple Silicon MPS available")
else:
print("CPU only — inference will be slower but still works")
# Always set this for Ampere+ GPUs (A100, RTX 3090, etc.)
torch.set_float32_matmul_precision("high")
Batch Size Tuning
# Start conservative, increase until OOM# GPU with 8 GB VRAM: per_core_batch_size=64# GPU with 16 GB VRAM: per_core_batch_size=128# GPU with 24 GB VRAM: per_core_batch_size=256# CPU with 8 GB RAM: per_core_batch_size=8# CPU with 16 GB RAM: per_core_batch_size=32# CPU with 32 GB RAM: per_core_batch_size=64
model.compile(timesfm.ForecastConfig(
max_context=1024,
max_horizon=256,
per_core_batch_size=32, # <-- tune this
normalize_inputs=True,
use_continuous_quantile_head=True,
fix_quantile_crossing=True,
))
Memory-Constrained Environments
import gc, torch
# Force garbage collection before loading
gc.collect()
if torch.cuda.is_available():
torch.cuda.empty_cache()
# Load model
model = timesfm.TimesFM_2p5_200M_torch.from_pretrained(
"google/timesfm-2.5-200m-pytorch"
)
# Use small batch size on low-memory machines
model.compile(timesfm.ForecastConfig(
max_context=512, # Reduce context if needed
max_horizon=128, # Reduce horizon if needed
per_core_batch_size=4, # Small batches
normalize_inputs=True,
use_continuous_quantile_head=True,
fix_quantile_crossing=True,
))
# Process series in chunks to avoid OOM
CHUNK = 50
all_results = []
for i inrange(0, len(inputs), CHUNK):
chunk = inputs[i:i+CHUNK]
p, q = model.forecast(horizon=H, inputs=chunk)
all_results.append((p, q))
gc.collect() # Clean up between chunks
🔗 Integration with Other Skills
With statsmodels
Use statsmodels for classical models (ARIMA, SARIMAX) as a comparison baseline:
CSV has 108 rows (3 stores x 36 weeks); stores have distinct price arrays
Quality Checklist
Run this checklist after every TimesFM task before declaring success:
Output shape correct -- point_fc shape is (n_series, horizon), quant_fc is (n_series, horizon, 10)
Quantile indices -- index 0 = mean, 1 = q10, 2 = q20 ... 9 = q90. NOT 0 = q0, 1 = q10.
Frequency flag -- TimesFM 1.0/2.0: pass freq=[0] for monthly data. TimesFM 2.5: no freq flag.
Series length -- context must be >= 32 data points (model minimum). Warn if shorter.
No NaN -- np.isnan(point_fc).any() should be False. Check input series for gaps first.
Visualization axes -- if multiple panels share data, use sharex=True. All time axes must cover the same span.
Binary outputs in Git LFS -- PNG and GIF files must be tracked via .gitattributes (repo root already configured).
No large datasets committed -- any real dataset > 1 MB should be downloaded to tempfile.mkdtemp() and annotated in code.
matplotlib.use('Agg') -- must appear before any pyplot import when running headless.
infer_is_positive -- set False for temperature anomalies, financial returns, or any series that can be negative.
Common Mistakes
These bugs have appeared in this skill's examples. Learn from them:
Quantile index off-by-one -- The most common mistake. quant_fc[..., 0] is the mean, not q0. q10 = index 1, q90 = index 9. Always define named constants: IDX_Q10, IDX_Q20, IDX_Q80, IDX_Q90 = 1, 2, 8, 9.
Variable shadowing in comprehensions -- If you build per-series covariate dicts inside a loop, do NOT use the loop variable as the comprehension variable. Accumulate into separate dict[str, ndarray] outside the loop, then assign.
# WRONG -- outer `store_id` gets shadowed:
covariates = {store_id: arr[store_id] for store_id in stores} # inside outer loop over store_id# CORRECT -- use a different name or accumulate beforehand:
prices_by_store: dict[str, np.ndarray] = {}
for store_id, config in stores.items():
prices_by_store[store_id] = compute_price(config)
Wrong CSV column name -- The global-temperature CSV uses anomaly_c, not anomaly. Always print(df.columns) before accessing.
tight_layout() warning with sharex=True -- Harmless; suppress with plt.tight_layout(rect=[0, 0, 1, 0.97]) or ignore.
TimesFM 2.5 required for forecast_with_covariates() -- TimesFM 1.0 does NOT have this method. Install pip install timesfm[xreg] and use checkpoint google/timesfm-2.5-200m-pytorch.
Future covariates must span the full horizon -- Dynamic covariates (price, promotions, holidays) must have values for BOTH the context AND the forecast horizon. You cannot pass context-only arrays.
Anomaly thresholds must be defined once -- Define CRITICAL_Z = 3.0, WARNING_Z = 2.0 as module-level constants. Never hardcode 3 or 2 inline.
Context anomaly detection uses residuals, not raw values -- Always detrend first ( linear, or seasonal decomposition), then Z-score the residuals. Raw-value Z-scores are misleading on trending data.
Validation & Verification
Use the example outputs as regression baselines. If you change forecasting logic, verify:
# Anomaly detection regression check:
python -c "
import json
d = json.load(open('examples/anomaly-detection/output/anomaly_detection.json'))
ctx = d['context_summary']
assert ctx['critical'] >= 1, 'Sep 2023 must be CRITICAL'
assert any(r['date'] == '2023-09' and r['severity'] == 'CRITICAL'
for r in d['context_detections']), 'Sep 2023 not found'
print('Anomaly detection regression: PASS')"# Covariates regression check:
python -c "
import pandas as pd
df = pd.read_csv('examples/covariates-forecasting/output/sales_with_covariates.csv')
assert len(df) == 108, f'Expected 108 rows, got {len(df)}'
prices = df.groupby('store_id')['price'].mean()
assert prices['store_A'] > prices['store_B'] > prices['store_C'], 'Store price ordering wrong'
print('Covariates regression: PASS')"
