Need Tiers as Long-Horizon Anchors for Human Activity Forecasting

Forecasts of what a person will do get worse the further ahead you look. For human activity sequences, such as check-ins, purchases, and trips, this decay is severe enough that on long horizons simple historical averages match or beat neural forecasters in most tasks. This proposal asks whether the decay is uniform across kinds of behavior. Psychology's needs hierarchies (Maslow, Alderfer) sort behavior into tiers. At the bottom are recurring bodily and safety needs such as eating, sleeping, commuting, and medical care. Above them sit social and self-directed pursuits. Lower-tier needs are homeostatic: they build up and must be discharged on a roughly fixed budget. That suggests they should stay predictable far longer than higher-tier behavior.

The core claim is that a forecaster should use this structure directly. At long lead times it would predict the total satisfaction each need tier will receive over a window, treating the bottom tiers as near-conserved budgets. It would produce individual events only over short windows, conditioned on those budgets. If the claim holds, it gives a principled way to decide what can still be forecast at day 14 and what cannot. It would also reconcile a standing tension. Reactive step-by-step generators drift, while rigid multi-step planners break when needs fluctuate. We plan the slow quantity (tier budgets) and react on the fast one (next event).

What the literature already shows

Two findings motivate this work. The first is the long-horizon "next-event trap". HoTPP Benchmark evaluates forecasters of timestamped labeled event streams (marked temporal point processes). It reports that models trained on one-step likelihood drift and collapse when unrolled, so that statistical baselines match or exceed neural models in 60% of long-horizon tasks, and that sequence-level decoders (diffusion, reranking) help. The second finding is that latent needs improve activity modeling. SAND represents physiological, safety and social needs as a latent vector evolving by a neural stochastic differential equation, meaning a learned continuous drift punctuated by jumps when an activity satisfies a need. It trains a policy by adversarial imitation and reports more than 20% and more than 40% reductions in Jensen-Shannon divergence on some sequence-statistic metrics. SAND, however, evaluates aggregate sequence fidelity and next-step prediction. It never measures how error grows with lead time, and it rolls out step by step, which is exactly the regime HoTPP identifies as fragile.

Other work constrains the design. D2A drives LLM agents with decaying desire scores. It found that adding a rigid multi-step planning module hurt realism, which argues against planning at the level of individual events. What Limits LLM-based Human Simulation reports GPT-4 human-likeness falling from 51% to 13.3% after 20 turns. This warns against relying on an LLM rollout as the forecaster.

Two further papers supply methods we reuse. Pose Transformers removes compounding error in motion forecasting with a non-autoregressive decoder anchored on a "copy the last pose" prior. Benchmarking Physics-Informed Time-Series Models shows that weather forecasts stay usable past 72 hours only when domain invariants are enforced as losses. Our homeostatic budget constraint plays the role those physical invariants play for weather.

Finally, ForecastBench-Sim documents error growth across seven horizons, but its agents follow game rules rather than psychological structure. The COVID studies (population-scale search needs, Nantong taxi trajectories) show that shocks compress behavior toward lower tiers. We use them only for a secondary test of prepotency, the psychological claim that lower-tier needs must be met before higher-tier ones drive behavior.

Core Claim

Forecast skill decays at tier-dependent rates, with lower tiers decaying slower. A model that forecasts tier budgets non-autoregressively and constrains lower-tier budgets to learned per-user homeostatic bands will beat both statistical baselines and autoregressive need models at long horizons. The gain should come from the tier structure and the constraint, not merely from coarsening labels or from non-autoregressive decoding.

To make "decays slower" testable, we define horizon-indexed skill against a reference forecaster:

$$S_k(H) = 1 - \frac{E_k^{\text{model}}(H)}{E_k^{\text{ref}}(H)},$$

Here $E_k(H)$ is the forecast error for tier $k$ over the window starting $H$ days ahead, and the reference is the per-user weekly historical profile. The tier's skill half-life $H^{1/2}_k$ is the smallest $H$ at which $S_k(H)$ falls to half its one-day value. The hypothesis is $H^{1/2}_{\text{physiological}} > H^{1/2}_{\text{safety}} > H^{1/2}_{\text{social}}$.

Research Program

The first part of the program is measurement. We build a horizon-decay protocol for three activity datasets. Two of them, Foursquare-NYC with 2,000 users, 14 labels and one month of data, and the Mobile Operator data with 10,000 users and 15 labels, already come with SAND's mapping of activities to need tiers. The third is HoTPP's Transactions stream with 203 merchant classes. We map it to tiers ourselves, for example grocery and pharmacy to physiological/safety and entertainment to higher tiers. Two annotators will do the mapping, and we will report their agreement. Because the check-in data spans only a month, horizons run from 1 to 14 days with a fixed 14-day lookback. Transactions allows longer horizons. We compute $S_k(H)$ and $H^{1/2}_k$ for every baseline, which by itself fills the unmeasured gap SAND leaves open.

The second part is the model. An encoder reads the lookback history. A non-autoregressive head predicts, for every future day up to the horizon, the tier-budget vector $b_{k,d}$: the need satisfaction delivered to tier $k$ on day $d$. The head is anchored, POTR-style, on the user's weekly profile, so it predicts deviations from that profile. A short-window event decoder then generates individual activities for a day, conditioned on that day's budgets. The homeostatic constraint is a penalty. It keeps each lower-tier budget inside a per-user band $[\ell_k, u_k]$ estimated from the lookback:

$$\mathcal{L} = \mathcal{L}_{\text{events}} + \lambda \sum_{k \in \text{lower}} \sum_d \big[\max(0, \ell_k - b_{k,d}) + \max(0, b_{k,d} - u_k)\big],$$

where $\lambda$ weights the constraint. Upper tiers receive no band, which encodes the prediction that they are not conserved.

The third part is a secondary, and admittedly weaker, test of prepotency under shocks. We use the Nantong trajectory data if we can obtain access; the file does not state whether it is released. We give the model a lockdown indicator and ask whether an ordering prior, that higher-tier budgets are cut before lower-tier ones, correctly predicts the direction of compression. The paper reports that the commuting share rose from 43.8% to 45.7% and medical travel from 7.1% to 8.1%. This test checks direction, not magnitude.

Baselines

The naive reference point is the per-user weekly historical profile. It predicts each future day's activity distribution as the average of the same weekday in the lookback. HoTPP's result makes this the forecaster to fear at long horizons. It needs no tuning, and it is also the denominator of $S_k(H)$. Beating it means positive skill at $H=14$ on at least the two lower tiers with a paired bootstrap confidence interval excluding zero. There is no reported number on Lacuna for this baseline on these datasets.

The strongest existing need-based method is SAND, retrained with its released setup, conditioned on each user's lookback, and rolled out autoregressively to every horizon. We report both the JSD sequence statistics SAND uses and our horizon metrics. The only reported figure on Lacuna is its relative claim of more than 20% and more than 40% JSD reductions over its baselines on some metrics. No absolute per-metric values were found. Beating SAND means lower tier-budget error and higher Temporal mean Average Precision (T-mAP, HoTPP's metric that scores predicted events by label and timing) at $H \geq 7$ while staying within its JSD at $H=1$.

The strongest need-free long-horizon method is HoTPP's best sequence-level decoder (diffusion or reranking) using the benchmark's code. We found no reported per-dataset T-mAP on Lacuna. We also run Dish-TS applied to the daily tier-count series, as a generic fix for the shift between lookback and forecast windows.

Every method gets the same lookback, the same splits by user and time, and the same budget of 40 hyperparameter trials selected on a validation split. Model sizes are matched to within 20% of parameters.

Three ablations isolate the key ingredient. The first replaces the Maslow tiers with an equal number of groups found by clustering activity co-occurrence, and separately with random groups. The second removes the homeostatic penalty. The third makes the budget head autoregressive.

Why not just…?

Why not just use the weekly historical profile? This is a real threat. HoTPP's 60% finding says it often wins, and very regular users may leave nothing to gain. Our claim is narrower: the model should beat it on users whose routines shift within the lookback, and on higher-tier behavior near $H=1$. We will stratify users by routine regularity and report where the profile remains unbeaten, rather than averaging the result away.

Why not just forecast at a coarser label granularity with any grouping? This is the most dangerous alternative. Much of the long-horizon benefit could come from aggregation alone. The clustered- and random-grouping ablations settle it. If learned clusters match Maslow tiers, the result becomes "aggregate first, forecast budgets" and the psychological framing is dropped. The homeostatic constraint can only be justified if it improves over the unconstrained version under the tier grouping specifically.

Why not just use a non-autoregressive or diffusion decoder, as HoTPP suggests? Sequence-level decoding may remove most compounding error by itself. We run exactly that decoder with and without the budget head, so the difference measures what the needs structure adds beyond the choice of decoding scheme.

Why not just apply Dish-TS normalization? Dish-TS predicts a future level and scale from statistics alone, and it reports over 10% gains over reversible instance normalization on long horizons. Applied to tier counts, it is a fair test of whether the band constraint does anything a learned rescaling cannot. If it matches our model, the invariant is redundant.

Proposed Experiments and Decisive Checks

The first check is the decay ordering. We look for $H^{1/2}_k$ ordered by tier, with non-overlapping bootstrap intervals, on at least two of the three datasets. A flat or reversed ordering falsifies the premise, and we would report it as the main result.

The second check is the long-horizon gain. The full model must beat the weekly profile, SAND and the HoTPP decoder at $H \geq 7$ on tier-budget error and T-mAP.

The third check is attribution. The Maslow grouping must beat clustered and random groupings, and removing the constraint must hurt only the lower tiers. If the constraint also hurts upper tiers, it is acting as generic regularization rather than as a needs invariant.

The shock-direction test on Nantong is reported only as supporting evidence.