Published: March 18, 2026 • Part of the Philippine Food Price Prediction series
This post documents the original system as published on March 5, 2026 — before the March 18 upgrades. It serves as a historical record of the baseline capabilities. For what changed, see the upgrade changelog. For the current system, see the upgraded system post.
System Overview (March 5, 2026)
The original Philippine Food Price Prediction system was a straightforward machine learning analysis of 153,404 WFP price observations spanning 2000–2023, covering 17 Philippine regions and 5 commodity categories (rice, corn, fish, meat, vegetables).
Model Architecture
| Component | Specification |
|---|---|
| Primary Model | Random Forest — 500 trees, max_depth=None (unlimited), min_samples_split=2 |
| Neural Network | Simple PyTorch feedforward — 2 hidden layers (50 neurons each), 4 input features, 100 epochs, CrossEntropyLoss |
| Training Data | 153,404 WFP price observations (2000–2023) |
| Features | Year, month, region (encoded), commodity (encoded) — 4 features total |
| Validation | Hold-out test set (2020–2023), R² = 0.91 |
| Ensemble | None — single model predictions only |
| External Data | None — WFP price data only, no exogenous features |
What Was Included
- Random Forest regression with feature importance analysis (lagged prices contributed 42% of predictive power)
- Simple feedforward neural network for classification (4 inputs → 2×50 hidden → output)
- Static HTML dashboard with Chart.js visualizations (light theme only)
- Regional price analysis — 17 Philippine regions with commodity breakdowns
- Temporal trend analysis — 23-year price trajectories per commodity
- Feature importance ranking — lagged prices, month, year, region
What Was NOT Included
These capabilities were listed as “Future Work” in the original post:
| Capability | Status (March 5) |
|---|---|
| LSTM deep learning model | ✗ Not implemented |
| Exogenous features (climate, exchange rates, food indices) | ✗ Not implemented |
| Climate scenario analysis (ENSO correlation) | ✗ Not implemented |
| Early warning system | ✗ Not implemented |
| Ensemble stacking model | ✗ Not implemented |
| Automated test suite | ✗ Not implemented (0 tests) |
| Dark mode / PWA / Export | ✗ Not implemented |
| REST API | ✗ Not implemented |
| Parallel training pipeline | ✗ Not implemented |
File Inventory (March 5)
| File | Lines | Purpose |
|---|---|---|
| retrain_model.py | ~350 | Random Forest training + predictions |
| daily_update.py | ~200 | WFP data download + model retraining |
| index.html | ~800 | Dashboard (Chart.js, light theme only) |
| README.md | ~50 | Basic setup instructions |
Key Findings (Unchanged)
- Rice prices increased 131.7% over 23 years (&peso;17.87 to &peso;41.41/kg)
- Two structural step-changes: 2008 global food crisis and 2014–2018 domestic supply tightening
- Regional disparities up to &peso;10/kg for rice (ARMM vs. Cagayan Valley)
- Lagged price features contribute 42% of Random Forest predictive power
- African Swine Fever crisis (2019–2021) produced 36.6% pork price surge
This was the baseline. See what changed → Upgrade Changelog | The Upgraded System
