Architecture Overview · Park Whisperer Platform

Attraction Wait Time Forecasting — LightGBM Iteration

LightGBM-based recursive multi-step forecasting for attraction wait times — one global model across all rides, predicting 7 horizons (15m to 3hr) by feeding predictions back as lag features. Abandoned at v6 due to recursive error accumulation.

LightGBM · GBDT · 7 HorizonsAbandoned · Informed v8 Design6-Hr Lookback Window

Click any component to explore it in the full case study.

Input A
Live PostgreSQL Wait Buffer
park_live_status_logs_v2 — 6-hour lookback per entity. JSONB path extraction from queue_details. Foundation for lag features and rolling statistics.
PostgreSQL · 6-hr lookback
Input B
Park Schedule Context
Today/yesterday/tomorrow operating hours from park_schedules_v2. Park open/close times used as features and for output clamping.
park_schedules_v2
Input C
METAR Weather (KISM)
KISM station: temperature, humidity, precipitation, wind speed/direction, pressure, sky cover, visibility. All normalized and joined to feature vector.
KISM · Hourly weather
LightGBM Recursive Multi-Step Forecasting
Step 1
Feature Assembly
Time features, lag (15m/30m/1hr/24hr/7day), rolling (1hr/4hr mean/std/max/min), weather, park schedule context. One global model for all attractions.
›
Step 2
Predict Horizon 1
LightGBM predicts 15-min ahead. max(0, predicted) clamp. Prediction appended to live buffer as if observed.
›
Step 3
Recursive Steps 2–7
Each predicted value becomes the lag input for the next step. 30m, 60m, 90m, 120m, 150m, 180m predictions built on prior predictions.
›
Step 4
Output Dict
{prediction_datetime → wait_seconds} for 7 horizons. One dict per entity.
⚠️ Why This Was Abandoned (Lesson)
Recursive lag imputation caused compounding error. By the 3-hour horizon, the buffer was entirely model-generated — not observed data. No real lag features at inference. Error accumulated multiplicatively. v8 (Prophet per-ride models) addressed this by eliminating lag features entirely and using temporal decomposition instead.
['Error accumulation at 3hr horizon', 'Buffer = model-generated by step 5+', 'No real lag at inference time', 'Single global model (all attractions)', 'Informed design of v8 (Prophet per-ride)', 'Lesson: recursive imputation is a trap']
ride-whisperer-v6.html#recursive forecasting
1
Global model (all attractions)
7
Forecast horizons
6hr
Lookback window
v6
Iteration
StackLightGBM (GBDT)PostgreSQL Flexible ServerPythonThreadPoolExecutorpsycopg2Pickle model artifactsKISM METAR