Absence Prediction Error: Rostering vs Neuroscience

Updated 3 July 2026

Absence prediction error is defined in rostering as forecast misclassification using confusion matrix metrics (α, β) to optimize reserve shift allocation.
In predictive-coding neuroscience, it denotes the discrepancy in Bernoulli belief updates when an expected sensory event is omitted.
The concept unifies operational decision-making and neural processing, emphasizing domain-specific error structures and implications for system design.

Absence prediction error is used in recent technical literature in two distinct senses. In workforce analytics and robust personnel rostering, it denotes the error structure of absenteeism forecasts that feed a predict-then-optimize pipeline, typically parameterized by the true-positive rate $\alpha$ and true-negative rate $\beta$, with downstream consequences for reserve-shift allocation and rerostering cost [2406.18119]. In predictive-coding neuroscience, it denotes a dedicated error signal that reduces a Bernoulli belief in stimulus occurrence when an expected sensory event is omitted; in that setting, absence prediction error is formalized as the current probability of stimulus presence at the moment of omission, $aPE_t = P_t(\mathrm{presence})$ [2511.21605]. The shared label masks a substantive difference: one usage concerns forecast misclassification in operational decision-making, whereas the other concerns belief updating about the non-occurrence of an expected event.

1. Terminological scope and formal contrasts

In the rostering literature, absence prediction error is not introduced as a single scalar residual. It is decomposed through the binary-classification confusion matrix. The relevant quantities are $TP$, $FN$, $FP$, and $TN$, from which the paper defines the true-positive rate $\alpha = TP/(TP+FN)$, the true-negative rate $\beta = TN/(TN+FP)$, the false-positive rate $FPR = 1-\beta$, and the false-negative rate $FNR = 1-\alpha$ [2406.18119]. Because the absenteeism process is strongly imbalanced, with actual absenteeism frequency $\rho = P(y_i=1)$, the same work rescales the false-positive rate and uses $rFPR = \rho \cdot FPR$.

In the auditory predictive-coding literature, by contrast, absence prediction error is defined over a binary belief about occurrence rather than over classifier outcomes. The prior is Bernoulli: $p = P(\mathrm{presence})$ and $1-p = P(\mathrm{absence})$. When no tone arrives at an expected time point, the relevant error is not a mismatch in a continuous feature value but the reduction of the belief in presence. Tabas et al. define $aPE_t = P_t(\mathrm{presence})$, and after omission the posterior in presence becomes zero, so the belief update is $\Delta P(\mathrm{presence}) = -aPE_t$ [2511.21605].

This contrast is structurally important. In the first usage, “error” is a property of a predictor evaluated against realized absenteeism; in the second, “error” is itself the mechanism by which perceptual belief is updated. This suggests that the term is application-specific rather than domain-general.

2. Predict-then-optimize rostering and the operationalization of absenteeism error

The rostering formulation begins with a binary absence classifier over employee-day pairs. For each employee-day $i$, the realized absence indicator is $y_i$, the predicted label is $\hat y_i$, and the class imbalance is encoded by $\rho = P(y_i=1)$. To avoid unrealistically many false positives when $\beta$ is low, the simulation uses $P(\hat y_i=1 \mid y_i=0) = rFPR = \rho \cdot (1-\beta)$ rather than the unscaled $FPR$ [2406.18119].

Predictions enter the optimization only after aggregation. For day $d$, the number of predicted absences is
$$
c_d^* = \sum_{n \in N} \hat y_n{}d.
$$
That daily quantity is then imposed as a reserve-shift requirement inside a robust rostering mixed-integer program. Once $c_d^*$ is fixed, the model enforces at least $c_d^*$ reserve shifts on each day. In the paper’s notation, with shift $s'$ denoting the reserve shift, the objective is
$$
\min \sum{n,d,s \neq s',k} \omega^i_n x_{n,d,s,k}
+ \sum_n \sum_{d,k} x_{n,d,s',k}\cdot \omega^7_n
+ \sum_d v^r_d \cdot \omega^r
$$
subject to demand, contractual, and reserve-buffer constraints [2406.18119].

The same setup can be expressed as a two-stage problem,
$$
\min_{x \in X} c(x) + E_{\hat y}[R(x,\hat y)],
$$
where $c(x)$ is the real-time wage cost plus the cost of scheduled reserves, and $R(x,\hat y)$ is the expected rerostering cost once true absences $y$ appear. In this formulation, absence prediction error propagates through $c_d^*$ into the first-stage reserve buffer. A plausible implication is that the operational value of prediction quality depends less on abstract classification accuracy than on how the induced reserve allocation interacts with skill substitutability and rerostering penalties.

3. Simulated machine learning, reserve buffers, and performance thresholds

Rather than training and testing actual machine-learning models, the rostering study simulates predictions at arbitrary performance levels. For each employee-day, it first draws the true outcome $y_i \sim \mathrm{Bernoulli}(\rho)$. If $y_i=1$, it draws $\hat y_i=1$ with probability $\alpha$; if $y_i=0$, it draws $\hat y_i=1$ with probability $rFPR$. The paper summarizes this as
python for each day d and employee n: yₙ_d ← Bernoulli(ρ) if yₙ_d = 1 then ŷₙ_d ← Bernoulli(α) else ŷₙ_d ← Bernoulli(rFPR)
which guarantees the desired marginal error rates [2406.18119].

The computational study uses a 35-nurse, 4-week instance with $\rho \approx 2.64\%$. Under uniform skills, even low $\alpha$ values of approximately $0.1$ suffice to outperform a policy that always schedules exactly one reserve per day. To beat a two-reserves-per-day baseline, the required performance rises to $\alpha \approx 0.3$–$0.4$, regardless of $rFPR$. The paper also reports that high $rFPR$ can partly compensate for low $\alpha$, because extra reserves, even when mis-placed, still help cover absences. Under hierarchical skills, however, reserve shifts must be assigned to the correct skill type to be effective. In that case, even perfect prediction of the total number $c_d^*$ cannot beat a fixed four-reserves-per-day rule, because the predictor does not allocate reserves to the right skill classes. To outperform the one-reserve-per-day baseline, moderately good $\alpha,\beta$ values remain sufficient, but higher performance is needed for stronger baselines [2406.18119].

These findings delimit the meaning of absence prediction error in an optimization setting. Reducing false negatives increases the probability that reserve capacity exists where it is needed, but reducing false positives is not uniformly beneficial if additional reserves have residual coverage value. This suggests that the relevant loss landscape is shaped by operational asymmetry rather than by symmetric classification criteria.

4. Severe class imbalance in individual-level absenteeism prediction

A related line of work studies individual-level absenteeism prediction as a time-series classification problem under severe class imbalance. Li et al. define $\rho = N^{+/N^-$} as the ratio of majority present samples to minority absent samples, with $\rho \approx 42$ in their study, and analyze Binary Focal Loss (BFL) and Geometric Mean (G-Mean) loss using only that imbalance ratio [2606.31532].

For BFL,
$$
FL(p,y)=
\begin{cases}
-\alpha (1-p)^\gamma \log p & \text{if } y=1,\
-(1-\alpha)p^\gamma \log(1-p) & \text{if } y=0,
\end{cases}
$$
and at initialization $p \approx 0.5$ the total gradient ratio satisfies
$$
|\nabla_{\mathrm{present}}|/|\nabla_{\mathrm{absent}}| \approx \rho \cdot \frac{\alpha}{1-\alpha}.
$$
Balanced gradients therefore require
$$
\alpha^* = \frac{1}{1+\rho},
$$
which gives $\alpha^* \approx 0.023$ when $\rho \approx 42$. For G-Mean loss,
$$
L_{GM} = 1 - \sqrt{\widehat{TPR}\cdot \widehat{TNR}},
$$
with
$$
\widehat{TPR} = \frac{1}{N^{+}\sum_{i:y_i=1}} p_i,\qquad
\widehat{TNR} = \frac{1}{N^{-}\sum_{j:y_j=0}(1-p_j).}
$$
Unlike BFL, G-Mean adapts automatically without parameter calibration [2606.31532].

The study formulates prediction as a temporal mapping from a past attendance window $X=(x_{t-n+1},\ldots,x_t)$ to future labels $y$ over the next $m$ days, rather than the contemporaneous mapping $X_t \to y_t$. It evaluates LSTM, CNN, and LSTM-FCN architectures. On LSTM-FCN with window size $80$, BFL with $\alpha^*=0.023$ and $\gamma=0$ achieves precision $0.995$, specificity $0.813$, and balanced accuracy $0.888$, while G-Mean yields precision $0.994$, specificity $0.844$, and balanced accuracy $0.734$. The LSTM-FCN also yields the highest precision and specificity among the architectures evaluated under G-Mean loss, and stable performance is reported for batch sizes $\ge 64$ and window sizes between $40$ and $80$ days [2606.31532].

This work does not define “absence prediction error” in the same sense as the rostering paper, but it clarifies how imbalance-aware loss design shapes the attainable false-positive and false-negative structure before any downstream optimization is applied.

5. Absence prediction error as a Bernoulli belief update in predictive coding

In the sensory-neuroscience usage, absence prediction error is introduced to explain how the brain updates beliefs not about stimulus features such as pitch, but about whether a stimulus occurred at all. Standard feature prediction error is described relative to a Gaussian prior with mean $\mu$ and variance $\sigma^2$, and is written as
$$
\delta_{\mathrm{feature}} \propto \frac{\mu_{\mathrm{observed}}-\mu_{\mathrm{predicted}}}{\sigma_{\mathrm{predicted}}^{2+\sigma_{\mathrm{sensory}}^2}.}
$$
Absence prediction error instead operates on a Bernoulli prior. If an expected tone is omitted, the relevant quantity is the predicted probability of presence at that time point:
$$
aPE_t = P_t(\mathrm{presence}).
$$
In the specific auditory paradigm, after hearing $k$ tones and knowing that an omission can occur only in positions $4$–$6$, the trial-by-trial belief is
$$
P_{k+1}(\mathrm{presence}) = (1-p_0) + p_0\cdot \left[1-\frac{1}{7-k}\right],
$$
with $p_0 = 6/7$ [2511.21605].

Participants heard trials of up to eight identical pure tones separated by $500$ ms. In $6/7$ of trials, exactly one tone was omitted in position $4$, $5$, or $6$; in the remaining $1/7$, there was no omission. Before each block, participants learned three rules: at most one omission per trial, some trials have no omission, and omissions can occur only in positions $4$–$6$, equally likely a priori. They then reported the omission’s position or indicated “no omission.” Within that design, $aPE$ is largest at position $4$ and smallest at position $6$, where the omission is fully expected and $P(\mathrm{presence})=1/7$. After the omission, the posterior belief in presence becomes zero, so the reduction of belief is exactly $-aPE_t$ [2511.21605].

This definition separates absence prediction error from ordinary omission responses. A mere entrainment account would predict identical BOLD responses for all omission positions, whereas the absence-prediction-error account predicts a graded response that scales inversely with the expectedness of omission.

6. Neural locus, related prediction-error concepts, and implications

The auditory study used high-resolution $1.5\ \mathrm{mm}^3$ 3 T fMRI targeted to the inferior colliculus (IC), medial geniculate body (MGB), and primary auditory cortex (PAC). A first-level GLM included six regressors: std0, std1, std2, om4, om5, and om6. Absence prediction error was tested with the contrast $BOLD(\mathrm{om4}) > BOLD(\mathrm{om6})$. All three ROIs showed repetition suppression for the adaptation contrast std0 $>$ std1. For omission responses, PAC was robust and MGB significant, whereas IC showed no reliable effect. For the graded absence-prediction-error contrast om4 $>$ om6, MGB showed a highly significant response, PAC a very strong response, and IC was centered on zero. After normalizing each ROI by its BOLD sensitivity, adaptation was strongest in IC, while omission response and absence prediction error were stronger in MGB than IC; there was no significant difference in absence-prediction-error strength between MGB and PAC [2511.21605].

The proposed interpretation is that feature prediction error appears already in IC and is relayed upward through MGB and cortex, whereas absence prediction error is absent in IC but emerges in MGB and persists in PAC. The authors therefore speculate about a distinct subcortical circuit, possibly involving dedicated thalamic neurons or microcircuits that compute a Bernoulli-type error for expected stimulus occurrence. They also suggest a possible “presence” counterpart that would exclusively increase presence probability, although that mechanism remains to be identified physiologically [2511.21605].

A useful terminological boundary follows from related motor-learning work. In force-field adaptation, removing visual error information suppresses visuomotor feedback gains by roughly $30$–$40\%$ in both early and late epochs, while leaving the pattern of gain modulation intact; that study concerns feedback tuning in the absence of visual error information, not absence prediction error as a belief-update signal about event occurrence [2306.00580]. This distinction matters because “absence” can refer either to a missing sensory feature cue or to the non-occurrence of an expected stimulus.

The broader implication drawn in the auditory study concerns psychopathology. Predictive-coding accounts of psychosis propose that hallucinations reflect a failure to down-weight erroneous predictions of stimulus presence when sensory input is absent. If the gain of the absence-prediction-error circuit is abnormally low, internally generated predictions of voices or tones would not be canceled by absence error. This suggests a mechanistic role for thalamic absence-prediction-error pathways in conditions where predictions about absence and presence are impaired [2511.21605].