what the stimulus explains

The movies repeat. A model that has learned what each neuron does at each frame can forecast well without capturing any dynamics of its own. These comparisons separate the two.

natural movie one

Classical
Neural
Stimulus-locked

One-step R² on this condition's test windows. Aqua rows are stimulus-locked references: the PSTH ignores neural history entirely.

natural movie three

Classical
Neural
Stimulus-locked

One-step R² on this condition's test windows. Aqua rows are stimulus-locked references: the PSTH ignores neural history entirely.

grey screen (spontaneous)

Classical
Neural
Stimulus-locked

No stimulus is on screen, so stimulus covariates carry only the fact that the screen is grey. Whatever history-only models achieve here is predictability without a stimulus to lock to.

the stimulus gain

The same model with and without the stimulus code, paired by session. Positive means the stimulus helped. Running speed is tested separately; a gain from running is a statement about prediction, not evidence that locomotion causes the activity.

History + stimulus minus history only (ΔR²)
Model · condition	Mean gain	95% CI	Sessions
VAR (ridge) · natural movie one	−0.0064	[−0.0126, −0.0009]	8
VAR (ridge) · natural movie three	−0.0098	[−0.0140, −0.0067]	8
VAR (ridge) · grey screen (spontaneous)	+0.0001	[−0.0009, +0.0010]	8
GRU · natural movie one	+0.0012	[−0.0099, +0.0176]	8
GRU · natural movie three	−0.0100	[−0.0146, −0.0053]	8
GRU · grey screen (spontaneous)	+0.0044	[−0.0006, +0.0117]	8
Transformer · natural movie one	−0.0042	[−0.0105, +0.0002]	8
Transformer · natural movie three	−0.0105	[−0.0174, −0.0056]	8
Transformer · grey screen (spontaneous)	−0.0035	[−0.0075, +0.0012]	8

gain by horizon

Classical
Neural

Stimulus gain in free rollout as the horizon grows. A gain that widens with horizon is the signature of a model leaning on the stimulus schedule rather than on the population's own recent past.