Adaptive signal systems such as the Sydney Coordinated Adaptive Traffic System (SCATS) controls traffic signal system by adjusting phase splits and cycle length in real time. The adjustments rely only on the last few seconds of detector data, the system lacks real-time predictive and optimization functionalities, allowing queues and emissions to build. Fine-tuning of signal timing by even a small increment can derive large economic and environmental benefits for the wider network. This short-coming is becoming increasingly significant in New Zealand, where transport already contributes 39 % of national CO₂ emissions and intersection delay costs Auckland more than NZ $1 billion each year. This thesis evaluates present and future performance of two representative SCATS-controlled intersections—Albany, Auckland and Ruakura, Hamilton—and tests whether supervised machine- learning (ML) models can predict delay and recommend cycle lengths that improvises the native SCATS logic. Field data was used to build, calibrate, and validate base SIDRA intersection models; saturation- flow rates were matched within ±5 % of observations and all lane-movements satisfied degree-of- saturation criteria. A Monte-Carlo routine expanded 14 SCATS peak-period volume logs into 98 synthetic volume scenarios, which were re-run in SIDRA to obtain delay, queue, fuel, and emission outputs. Four machine learning models—XGBoost, Random Forest, Support Vector Regression, and k-Nearest Neighbours—were trained on the synthetic dataset; the best two, XGBoost and Random Forest were combined in an ensemble to give a delay-prediction model. Furthermore, baseline analysis (2024) found both sites operating at Level-of-Service D, with several right-turn lanes already oversaturated and 95th-percentile queues exceeding storage. Ten- year growth projections degraded both intersections to LoS F well before the planning horizon. The ML ensemble predicted average control delay with MAE ≈ 4 s veh⁻¹ and, under moderate demand, shortened cycles by 15–46 s, cutting delay by 13 % and CO₂ by 5 % relative to SCATS timings; though the benefits diminished under heavy oversaturation. Data-driven machine learning models can provide cycle-by-cycle delay forecasts and substantial performance gains, but require additional high-degree of saturations and longer cycle lengths training data for robust operation and more generalization during severe congestion. Integration of such predictors and optimizers with SCATS would provide practical step toward meeting New Zealand’s 2035 emission-reduction targets.