Oil compressibility (co) plays a vital role in vast aspects ranging from upstream to downstream. For reservoir with pressure below bubble point, the effect of co to the fluid flow is insignificant as it is overshadowed by the presence of large gas compressibility (cg). This study aims to increase the range of applicability and accuracy of the formula used for estimating the co by eliminating the limitations of other existing correlations. A new formula for the estimation of oil compressibility below bubble point pressure is devised using Group Method Data Handling (GMDH). The approach is a combination of neural networks and some high-level statistical methods which rely on generating simple relations among the input parameters and the dependent parameter. The relations then result in eliminating some parameters with low impact on the output. A series of consecutive layers with the link is generated, and polynomial terms are created. A total number of 322 data points were collected from different sources from literature. Systematic trend analysis has been conducted to verify that the proposed GMDH model honours the exact physical behaviour. The new proposed model found to follow the correct trend, which implies its soundness. Besides, a comparative study was carried out using the best available correlations to confirm the significance of the results of the oil compressibility prediction using GMDH. Different statistical analyses have been conducted to verify the robustness of the newly developed model. The statistical analyses showed a positive outcome whereby the proposed model obtained the lowest average absolute percentage relative error of 5.17 and the highest correlation coefficient of 96.8. The best model tested among the other models has five input parameters and an average absolute percentage relative error of 10.955 and a correlation coefficient of 95.6. The new approach managed to reduce the curse of dimensionality as four input parameters have found to have a strong dependency on co (solution gas-oil ratio, oil density, reservoir temperature, and reservoir pressure). The new proposed model overcome the limitations described by the locality of some correlations as they depend on data from specific locations. Copyright 2019, Society of Petroleum Engineers.