Forecasting (internal-state prediction)

The third arc of the sensory-control-loop, named by Petzschner et al. (2021): the CNS predicts how internal states change not only as a consequence of actions, but also as a function of their own internal dynamics.

The second clause is what distinguishes this from anything else the wiki holds. Bodies do things on their own — glucose falls, temperature drifts, inflammation runs a course — and any anticipatory regulation has to model that trajectory, not just the effect of its own interventions.

The empirical hook

The paper’s example is Zimmerman et al.’s thirst neurons: drinking terminates minutes before any appreciable change in blood osmolarity. The satiation signal cannot be a reading of the regulated variable, because the regulated variable has not moved yet. Something is running the body forward in time and reporting the predicted outcome of the action already taken. That is forecasting in the strict sense, in a genetically tractable circuit — and the cleanest existence proof the wiki has that internal-state prediction is a real computation rather than a theoretical convenience.

Two directions to run it

• Forward (action → outcome): simulate the future sensory consequences of an action, plus the intrinsic dynamics of the body. A model of what will happen.
• Backward (outcome → action): assume a desired future state and simulate which actions would reach it, selecting the most promising. This makes forecasting part of action selection rather than a separate faculty — and blurs the boundary with the second arc of the loop.

Why this page is mostly a gap

Both major frameworks of body regulation assume forecasting without formalizing it. homeostatic-reinforcement-learning needs it (to predict which action reduces drive); active-inference needs it (its descending predictions are about future states). But the paper’s verdict is that “there are few frameworks that have formalized forecasting, especially in the internal domain” — the citations it can offer are Sutton & Barto’s RL, Botvinick & Toussaint’s planning-as-inference, and one dynamic Bayesian model of homeostatic control (Penny & Stephan, 2014).

The wiki’s own shelf is emptier still. Every source here that touches anticipation touches it at the level of what anticipatory regulation accomplishes — allostasis in Sterling’s and Barrett’s senses, Paulus’s “failure to anticipate interoceptive states” as the seat of anxiety and depression, Craig’s gesture at the AIC representing “predictions of future feelings.” None of them models the trajectory being predicted.

That matters for a claim the wiki leans on hard. interoceptive-psychopathology and computational-psychiatry locate psychopathology in mis-set predictions about the body — but a mis-set prediction about a future bodily state presupposes a model of how bodily states unfold, and no one has written that model down for the internal domain. The clinical account is currently resting on the least-formalized arc of the loop.

The near miss, and why it is worth recording as a miss

Harrison et al. (2021) fit a learning model trial by trial to predictions about upcoming inspiratory resistance, and get neural correlates for both prediction certainty and prediction error (BLT). It is the closest thing the wiki has to an interoceptive prediction being measured rather than postulated, and it is tempting to file it here.

It does not belong here. What their participants learn is that an arbitrary visual cue predicts an externally imposed event — ordinary associative learning whose outcome happens to be interoceptive. The body does nothing under its own dynamics; the experimenter applies a resistance or does not. Nothing in the design models a trajectory of internal state.

The distinction is the one this page was created to hold, and the near miss makes it sharper rather than blurring it. Predicting that a stimulus is coming is not predicting how a bodily state will evolve. Zimmerman’s thirst neurons remain the wiki’s only existence proof of the second, and they are in mice.

So the gap narrows in one respect and not at all in the other: interoceptive prediction is now measurable in humans, and internal-state forecasting still has no instrument, no model in the wiki’s literature, and — on computational-psychiatry’s account — a clinical framework resting on it.

The second near miss, from an unrelated lineage

Wilzok et al. (2023) arrive at the same boundary from the opposite direction — no learning model, no scanner, just asking people to rate what they expect their body to feel like — and land on the same side of it.

Their two paradigms measure anticipated versus experienced magnitude of pinprick pain and inspiratory loading, per trial, before the stimulus, in two channels, reliably over a week. The paper’s language is squarely this page’s: they set out to operationalize “the anticipation of potential future interoceptive states,” which they say “enables organisms to take regulatory actions to meet demands before they arise” — an allostatic framing, and a fair description of what forecasting is for.

What their participants actually anticipate is that a shade of grey will be followed by an experimenter applying a calibrated force to their forearm. The internal state has no dynamics in the design. Cue, wait, stimulus, rate.

Two near-misses now, from lineages that do not cite each other — Harrison’s from computational psychiatry, Wilzok’s from clinical psychosomatics — failing in the identical way. That repetition is itself informative, and it is why this page keeps the distinction rather than relaxing it. The available design for studying interoceptive prediction is cue-signals-imposed-event, because it is the design that gives the experimenter control of the outcome. Forecasting requires the opposite: letting a bodily state evolve on its own and asking what the person expected it to do. Nobody has built that experiment for humans, and the reason may be that the obvious paradigm family structurally cannot contain it.

The terminological hazard is now real enough to name. Wilzok et al. call their measure the anticipation of future interoceptive states; Paulus & Stein’s clinical thesis is about the failure to anticipate interoceptive states; this page reserves forecasting for something neither of them measures. Anyone reading across those three will find the same phrase doing three jobs. See discrepancy-awareness for the three-way table that keeps them apart.

Relation to predictive homeostasis and allostasis

Keep three things apart:

• Predictive homeostasis — acting to move away from the set-point in anticipation of a perturbation (raising temperature before expected cold).
• allostasis — changing the set-point itself.
• Forecasting — the model of future internal states that both of the above require in order to be anticipatory at all.

The first two name regulatory strategies; this page names the predictive machinery they depend on. The wiki has substantial material on the strategies and effectively none on the machinery, which is the point of recording the distinction.