Background: In neonatal encephalopathy, the clinical manifestations of injury can only be reliably assessed several years after an intervention, complicating early prognostication and rendering trials of promising neuroprotectants slow and expensive. We aimed to determine the accuracy of thalamic proton magnetic resonance (MR) spectroscopy (MRS) biomarkers as early predictors of the neurodevelopmental abnormalities observed years after neonatal encephalopathy. Methods: We did a prospective multicentre cohort study across eight neonatal intensive care units in the UK and USA, recruiting term and near-term neonates who received therapeutic hypothermia for neonatal encephalopathy. We excluded infants with life-threatening congenital malformations, syndromic disorders, neurometabolic diseases, or any alternative diagnoses for encephalopathy that were apparent within 6 h of birth. We obtained T 1 -weighted, T 2 -weighted, and diffusion-weighted MRI and thalamic proton MRS 4–14 days after birth. Clinical neurodevelopmental tests were done 18–24 months later. The primary outcome was the association between MR biomarkers and an adverse neurodevelopmental outcome, defined as death or moderate or severe disability, measured using a multivariable prognostic model. We used receiver operating characteristic (ROC) curves to examine the prognostic accuracy of the individual biomarkers. This trial is registered with ClinicalTrials.gov, number NCT01309711. Findings: Between Jan 29, 2013, and June 25, 2016, we recruited 223 infants who all underwent MRI and MRS at a median age of 7 days (IQR 5–10), with 190 (85%) followed up for neurological examination at a median age of 23 months (20–25). Of those followed up, 31 (16%) had moderate or severe disability, including one death. Multiple logistic regression analysis could not be done because thalamic N-acetylaspartate (NAA) concentration alone accurately predicted an adverse neurodevelopmental outcome (area under the curve [AUC] of 0·99 [95% CI 0·94–1·00]; sensitivity 100% [74–100]; specificity 97% [90–100]; n=82); the models would not converge when any additional variable was examined. The AUC (95% CI) of clinical examination at 6 h (n=190) and at discharge (n=167) were 0·72 (0·65–0·78) and 0·60 (0·53–0·68), respectively, and the AUC of abnormal amplitude integrated EEG at 6 h (n=169) was 0·73 (0·65–0·79). On conventional MRI (n=190), cortical injury had an AUC of 0·67 (0·60–0·73), basal ganglia or thalamic injury had an AUC of 0·81 (0·75–0·87), and abnormal signal in the posterior limb of internal capsule (PLIC) had an AUC of 0·82 (0·76–0·87). Fractional anisotropy of PLIC (n=65) had an AUC of 0·82 (0·76–0·87). MRS metabolite peak-area ratios (n=160) of NAA–creatine (<1·29) had an AUC of 0·79 (0·72–0·85), of NAA–choline had an AUC of 0·74 (0·66–0·80), and of lactate–NAA (>0·22) had an AUC of 0·94 (0·89–0·97). Interpretation: Thalamic proton MRS measures acquired soon after birth in neonatal encephalopathy had the highest accuracy to predict neurdevelopment 2 years later. These methods could be applied to increase the power of neuroprotection trials while reducing their duration. Funding: National Institute for Health Research UK.