Esophageal squamous cell carcinoma (ESCC) is a common malignant gastrointestinal tumor threatening global human health. For patients diagnosed with ESCC, determining the prognosis is a huge challenge. Due to their important role in tumor progression, long non-coding RNAs (lncRNAs) may be putative molecular candidates in the survival prediction of ESCC patients. Here, we obtained three datasets of ESCC lncRNA expression profiles (GSE53624, GSE53622, and GSE53625) from the Gene Expression Omnibus (GEO) database. The method of statistics and machine learning including survival analysis and LASSO regression analysis were applied. We identified a six-lncRNA signature composed of AL445524.1, AC109439.2, LINC01273, AC015922.3, LINC00547, and PSPC1-AS2. Kaplan-Meier and Cox analyses were conducted, and the prognostic ability and predictive independence of the lncRNA signature were found in three ESCC datasets. In the entire set, time-dependent ROC curve analysis showed that the prediction accuracy of the lncRNA signature was remarkably greater than that of TNM stage. ROC and stratified analysis indicated that the combination of six-lncRNA signature with the TNM stage has the highest accuracy in subgrouping ESCC patients. Furthermore, experiments subsequently confirmed that one of the lncRNAs LINC01273 may play an oncogenic role in ESCC. This study suggested the six-lncRNA signature could be a valuable survival predictor for patients with ESCC and have potential to be an auxiliary biomarker of TNM stage to subdivide ESCC patients more accurately, which has important clinical significance.