Lumber Market Watch

Transparency

Every number has a source.
Here's all of them.

Lumber Market Watch is built entirely on publicly available government data feeds and open APIs. This page documents every data source, update frequency, and the exact fields used by the model — so you can verify or reproduce any output.

Data Sources

🌊
NOAA CO-OPS↗

Center for Operational Oceanographic Products and Services

Every 6 minutes (live readings)

What it is

Real-time and historical water level (tide gauge) data from over 200 coastal stations across the United States.

Why it matters

Coastal storm surge is the primary supply-chain disruption mechanism for the lumber industry. When water levels exceed 3 ft above MHHW at key East Coast ports, timber supply routes are compromised — reflected in elevated CME Lumber Futures prices 3–9 months later.

How it's used

We query 12 NOAA CO-OPS stations along the US East Coast (from Maine to Florida) every 6 hours via the NOAA API. Max predicted surge values are passed as features to the LSTM model.

Fields used

water_levelpredictedpreliminaryverifiedstation_iddate_time

Raw training file

historical_tide_surges.csv, historical_offshore_buoys.csv
📡
NOAA NDBC↗

National Data Buoy Center

Every hour (buoy telemetry)

What it is

Offshore wave height, wind speed, water temperature, and atmospheric pressure from a network of moored and drifting buoys.

Why it matters

Offshore buoy data captures storm intensity before it reaches the coast. High significant wave height (Hs > 3m) combined with strong onshore winds are leading indicators of surge events. Buoy data extends the model's predictive horizon by 6–18 hours.

How it's used

Selected offshore buoys in the Western Atlantic are polled hourly. Significant wave height (WVHT) and dominant wave period (DPD) are used as secondary features in the LSTM.

Fields used

WVHTDPDAPDWDIRWSPDPRESATMPWTMP

Raw training file

historical_offshore_buoys.csv
🌀
HURDAT2↗

Atlantic Hurricane Database (Re-analysis Project)

Annually (historical archive)

What it is

The complete historical record of all Atlantic tropical cyclones from 1851 to present, including track positions, maximum sustained winds, and central pressure at 6-hour intervals.

Why it matters

The LSTM model was trained on historical surge events that correlate with HURDAT2 storm tracks. This gives the model a physical understanding of how storm track proximity, intensity, and forward speed map to surge amplitude at specific stations.

How it's used

HURDAT2 provides the training labels for the classification task. We identify historical "high surge" events at each NOAA station and link them to cyclone tracks within 500 km. The model learns the spatial-temporal relationship between storm geometry and surge outcome.

Fields used

datetimerecord_idstatuslatlonmax_windmin_pressure

Raw training file

hurdat2_processed.csv, east_coast_hurdat2.csv
📈
Alpaca Markets↗

Alpaca Securities LLC (Paper Trading API)

Real-time (market hours)

What it is

Commission-free paper and live trading API for US equities. We use Alpaca to track portfolio performance, submit simulated trades, and stream real-time stock quotes.

Why it matters

Paper trading allows us to measure the model's financial performance without real capital at risk. The account dashboard reflects exactly what live performance would look like.

How it's used

When the model emits a BUY signal, the Alpaca API simulates a purchase of the ETF proxy basket (LPX, WY, RFP, PCH). Position sizing is proportional to model confidence. Trades are logged and P&L is tracked continuously.

Fields used

portfolio_valuecashpositionsordersday_plequity

Raw training file

N/A (live API)
📉
Macro Indicators↗

Financial Modeling Prep / Yahoo Finance

Daily (market close)

What it is

Real-time and end-of-day S&P 500 index values and US 10-Year Treasury Yield — the two macro inputs to the signal filter.

Why it matters

Even a perfect surge forecast cannot predict a lumber rally in a severe risk-off macro environment (e.g., 2008-style credit crisis). The model uses macro regime detection to suppress BUY signals when the equity market is in free-fall or credit conditions are tightening sharply.

How it's used

S&P 500 and 10-yr yield are fetched daily via FMP API. If the S&P 500 has declined > 15% from its 52-week high AND the 10-yr yield is spiking, the BUY signal is automatically suppressed regardless of surge conditions.

Fields used

SPY_closeTNX_closedate

Raw training file

Model_B_Financial_Training_Set.csv

Flood Parcel Impact by Region

For each NOAA monitoring station, the table below shows the number of parcels, estimated population, and assessed property value that fall within FEMA Special Flood Hazard Area (SFHA / 100-year flood zone) polygons. When the model predicts surge above the 3.0 ft threshold at a given station, these are the assets directly at risk of inundation. Data sourced from FEMA NFIP FIRM panels and state GIS flood zone databases.

ResidentialCommercialIndustrialOther
Eastport, MEMaine
FIRM 2020-05-19
Parcels at Risk680
Flood Zone Acres85 ac
Pop. at Risk920
Est. Property Value$42M
Breakdown
RES
63%
COM
19%
IND
13%
OTH
5%
Primary RiskQuoddy Head commercial fishing; marine supply chain via Passamaquoddy Bay
FEMA NFIP · ME DACF GIS
Portland, MEMaine
FIRM 2021-08-10
Parcels at Risk2,800
Flood Zone Acres340 ac
Pop. at Risk4,200
Est. Property Value$285M
Breakdown
RES
66%
COM
21%
IND
8%
OTH
6%
Primary RiskPort of Portland timber & paper imports; Back Cove tidal inundation
FEMA NFIP · ME DACF GIS
Sewells Point, VAVirginia
FIRM 2020-11-20
Parcels at Risk18,200
Flood Zone Acres2,100 ac
Pop. at Risk38,000
Est. Property Value$4.2B
Breakdown
RES
68%
COM
18%
IND
10%
OTH
4%
Primary RiskNorfolk / Hampton Roads shipping lanes; highest land-subsidence rate on US East Coast (4–5 mm/yr)
FEMA NFIP · VA DCR · HRSD
Wrightsville Beach, NCNorth Carolina
FIRM 2022-03-15
Parcels at Risk3,400
Flood Zone Acres410 ac
Pop. at Risk5,800
Est. Property Value$680M
Breakdown
RES
86%
COM
10%
IND
2%
OTH
2%
Primary RiskWilmington waterfront residential; Port of Wilmington lumber & panel transit
FEMA NFIP · NC DEMLR
Charleston, SCSouth Carolina
FIRM 2023-01-06
Parcels at Risk9,600
Flood Zone Acres1,200 ac
Pop. at Risk22,000
Est. Property Value$2.8B
Breakdown
RES
65%
COM
22%
IND
8%
OTH
6%
Primary RiskCharleston Harbor port operations; historic district tidal flooding; industrial waterfront
FEMA NFIP · SC DHEC · City of Charleston
Springmaid Pier, SCSouth Carolina
FIRM 2021-10-22
Parcels at Risk5,200
Flood Zone Acres620 ac
Pop. at Risk11,000
Est. Property Value$1.4B
Breakdown
RES
73%
COM
20%
IND
3%
OTH
3%
Primary RiskMyrtle Beach resort corridor; Grand Strand Intracoastal Waterway blockage
FEMA NFIP · SC DHEC · Horry County
Fernandina Beach, FLFlorida
FIRM 2022-07-29
Parcels at Risk4,100
Flood Zone Acres480 ac
Pop. at Risk7,200
Est. Property Value$890M
Breakdown
RES
68%
COM
19%
IND
8%
OTH
5%
Primary RiskAmelia Island waterfront; Port of Fernandina paper & pulp export terminal
FEMA NFIP · FL DEP · Nassau County
43,980Total parcels monitored
5,235 acTotal flood zone acres
89,120Total population at risk
$10.3BTotal property at risk

⚠ Parcel counts and property values are estimates derived from FEMA FIRM panels and county assessor data. Actual inundation extent depends on storm track, duration, and antecedent soil saturation. These figures inform the model's supply-chain disruption signal — not official FEMA flood insurance determinations.

Model Architecture

The signal is generated by two stacked models: an LSTM surge classifier and a macro-aware financial filter. Here's exactly what each component does.

Model type
LSTM (Long Short-Term Memory)
A type of recurrent neural network designed for time-series data. It learns temporal dependencies — e.g., "surge at station A is usually followed by surge at station B 6 hours later."
Training data range
2005 – 2023
18 years of surge events, covering major storms including Katrina, Sandy, Irma, and Ian.
Signal type
Binary classification (BUY / HOLD)
Output 1 = expect > 5% lumber price increase within 9 months. Output 0 = hold / no position.
Confidence score
Softmax probability [0–100%]
The raw model probability of the predicted class. Higher confidence = stronger signal. Signals below 55% confidence are treated as noise.
Lookback window
14 days
The model considers the prior 14 days of surge readings when making a prediction.
Financial classifier
Gradient Boosting (Financial Model B)
A secondary model that takes the LSTM surge score + current macro conditions and outputs the final trading signal. Prevents trading into hostile macro environments.

Signal Pipeline

🌊NOAA CO-OPSTide gauges
→
📡NDBC BuoysOffshore sensors
→
🧠LSTM Model14-day lookback
→
📉Macro FilterS&P + 10-yr yield
→
⚡BUY / HOLDDaily signal

⚠️ Important Disclaimer

Lumber Market Watch is a research and educational project. All signals are for informational purposes only and do not constitute financial advice. Past model performance does not guarantee future results. CME Lumber Futures are complex instruments with significant risk of loss. Always consult a licensed financial advisor before making trading decisions. The authors of this project are not registered investment advisors.