Résumé : Background Initial improvement in the first weeks of antidepressant (AD) treatment is a useful early predictor of complete AD response. We performed a meta-analysis of AD studies to investigate whether a partial decrease in depressive symptoms by week 4 was associated with response and remission by weeks 6–14 in major depressive disorder (MDD). Finally, we focused on treatment-resistant depression (TRD: lack of response to prior AD) to test the impact of early improvement on a second AD treatment outcome and to compare different switching strategies. Methods Meta-analysis was conducted on AD naturalistic studies published between 01.01.2000 and 06.30.2017. TRD was an exclusion criterion. TRD was analyzed in 407 MDD patients treated with venlafaxine for 6 weeks. The MADRS was used to define very early improvement (VEI: > 20% decrease at week 2), early improvement (EI: > 30% decrease at week 4) and remission (week 6 MADRS < 10). A theoretical model was used to simulate AD switch in TRD patients who failed to achieve remission (Algorithm A), VEI (Algorithm B) or EI (Algorithm C). Results Our meta-analysis (9 studies; N = 6185) showed significant associations between early improvement, response (OR: 3.28 95% C.I: 2.06–5.20) and remission (OR: 2.10 95% C.I: 1.53–2.87). 24.6% of TRD sample remitted. VEI was a poor outcome predictor: sensitivity = 0.52 (0.40–0.63); specificity = 0.82 (0.76–0.86); AUC = 0.67 (0.62–0.71). EI had a moderate predictive power: sensitivity = 0.87 (0.77–0.93); specificity = 0.71 (0.66–0.77); AUC = 0.76 (0.71–0.80). The best treatment scenario was Algorithm C (switch after 4 weeks) in which remission rate was marginally increased (35.1% vs 33.7% of Algorithm A). Algorithm B (switch after 2 weeks) led to a 4.3% decrease in remission compared to Algorithm A. Limitations Inclusion of a naturalistic sample without a control arm; simulation of treatments. Conclusion Although literature data suggest a correlation between an initial improvement of depressive symptoms and later response and remission during AD treatment, our analysis shows that such an early improvement is not a reliable outcome predictor in TRD. The nature of TRD is complex and different biological mechanisms and treatments might be necessary for TRD patients.