Intraday Reforecasting and Real-Time Forecast Updates

Intraday Reforecasting and Real-Time Forecast Updates covers the forecasting methodology for updating a demand forecast mid-day as actual data arrives. This is distinct from real-time adherence management — the operational response to variance. Reforecasting is about revising the statistical prediction itself, not about moving agents in response to a queue spike.

The fundamental insight: a forecast made three days ago using historical patterns is strictly less informed than a forecast updated at 10am using four hours of today's actuals. Reforecasting exploits this information advantage to produce tighter predictions for the remaining intervals, which in turn produce better staffing decisions.

The information gap

At the start of a day, the WFM forecast is a prior — the best estimate given historical patterns, day-of-week effects, trend, and any known events. As intervals complete, each actual observation carries new information:

• Is today running hotter or colder than expected?
• Is the deviation consistent across intervals or concentrated in specific periods?
• Does the pattern suggest an underlying driver (marketing event, system outage) or random noise?

A reforecast incorporates this information into the remaining-day prediction. The value depends on how much of the day has elapsed (more actuals = more information) and how correlated intervals are within a day (higher intraday autocorrelation = more value from early observations).

Method 1: Simple ratio adjustment

The simplest and most widely used method in WFM platforms. Compute the ratio of actual-to-forecast for completed intervals, then apply that ratio to the remaining forecast.

Calculation

Let $F_t$ be the original forecast and $A_t$ the actual for interval $t$. At the reforecast point $k$ (after $k$ intervals have completed):

$\text{Ratio}=\frac{\sum _{t=1}^k{A}_t}{\sum _{t=1}^k{F}_t}$

For each remaining interval $t>k$:

$F{\text{'}}_t=F_t\times \text{Ratio}$

If the morning ran 12% hot (ratio = 1.12), every remaining interval gets multiplied by 1.12.

Strengths

• Trivially simple to implement and explain
• Available in virtually every WFM platform as a built-in feature
• Works well when the deviation is uniform across the day — a "level shift" where the entire day is proportionally hotter or colder than expected

Weaknesses

• Assumes the deviation pattern is constant. If the morning spike is driven by a marketing email that will taper off by afternoon, the ratio method will over-adjust the afternoon intervals.
• A single outlier interval (system outage at 9am producing a spike of queued calls) distorts the ratio for the entire remaining day.
• Does not account for the intraday profile shape — the relationship between morning and afternoon volumes may not be purely multiplicative.

When to use

Ratio adjustment is the right default when the deviation appears to be a level shift (proportional and consistent), and when the practitioner has no reason to believe the afternoon pattern will differ from the morning pattern. Most routine reforecasting in WFM operations uses this method because it is fast, transparent, and good enough for the majority of days.

Method 2: Exponential smoothing update

A more sophisticated approach that blends the original forecast with actual observations using a smoothing parameter, giving more weight to recent actuals.

Calculation

At each completed interval, update the forecast adjustment factor using exponential smoothing:

${\hat{r}}_t=\alpha \cdot \frac{A_t}{F_t}+(1-\alpha )\cdot {\hat{r}}_{t-1}$

where $\alpha$ is the smoothing parameter (typically 0.1 to 0.3 for intraday updates) and ${\hat{r}}_0=1.0$ (the prior that the forecast is correct).

The remaining intervals are then adjusted:

$F{\text{'}}_t=F_t\times {\hat{r}}_k$

where ${\hat{r}}_k$ is the smoothed ratio at the most recent completed interval.

How alpha controls reactivity

• $\alpha =0.1$ — slow to react; needs many intervals of consistent deviation before the adjustment becomes large. Good for stable days where you want to avoid chasing noise.
• $\alpha =0.3$ — moderately reactive; picks up sustained deviations within a few intervals. The practical sweet spot for most WFM applications.
• $\alpha =0.5+$ — very reactive; recent intervals dominate. Approaches the behavior of using only the last interval's ratio. Useful during volatile events but risks overreaction.

Strengths

• Naturally dampens the impact of single-interval outliers (unlike the simple ratio, which includes them at full weight)
• Adapts the adjustment factor as new information arrives — if the morning deviation reverses in late morning, the smoothed ratio moves back toward 1.0
• The smoothing parameter provides an explicit control lever for sensitivity

Weaknesses

• Requires choosing alpha, which introduces a tuning decision. Most WFM teams default to alpha = 0.2 and never revisit it.
• Still assumes a multiplicative relationship between deviation and the remaining forecast profile

When to use

Exponential smoothing updates are preferable to simple ratios when the day is volatile, when early intervals include known anomalies (brief system outages, batch arrivals), or when the WFM team wants a more conservative adjustment that does not overreact to short-term noise. Several commercial WFM platforms implement this method behind an "adaptive reforecast" or "smart intraday" feature label.

Method 3: Bayesian posterior update

The most principled method, treating the original forecast as a prior distribution and the actual observations as likelihood data, then computing the posterior distribution for the remaining day.

Conceptual framework

• Prior: The original forecast for the day, expressed as a distribution. For interval-level volume, a common choice is a Normal distribution centered on the forecast with variance estimated from historical forecast errors: $F_t\sim N({\hat{y}}_t,{\sigma }_F^2)$.
• Likelihood: The probability of observing the actuals given a "true" daily volume level. As each interval completes, the likelihood tightens — more data constrains the plausible range of "true" volume.
• Posterior: The updated distribution for remaining-day volume, incorporating both the prior forecast and the actual observations.

Simplified implementation

For a Normal-Normal conjugate model (prior and likelihood both Normal), the posterior has a closed-form solution. Let the prior for total daily volume be $N({\mu }_0,{\sigma }_0^2)$ where ${\mu }_0$ is the original daily forecast and ${\sigma }_0^2$ is the historical forecast variance. After observing $k$ intervals with cumulative actual $A_{1:k}$:

${\mu }_{\text{post}}=\frac{{\sigma }_0^2\cdot A_{1:k}/p_k+{\sigma }_L^2\cdot {\mu }_0}{{\sigma }_0^2+{\sigma }_L^2}$

where $p_k$ is the proportion of daily volume expected in the first $k$ intervals (from the intraday profile) and ${\sigma }_L^2$ is the observation variance.

The posterior mean is a precision-weighted blend of the prior forecast and the actual-implied daily total. Early in the day (few observations, wide likelihood), the posterior stays close to the prior. Late in the day (many observations, tight likelihood), the posterior converges on the actual-implied total.

Strengths

• Provides a full posterior distribution, not just a point reforecast — enabling probabilistic staffing decisions for the remainder of the day
• Naturally calibrates the weight given to actuals vs. the original forecast based on how much data has arrived
• Formally handles the transition from "mostly prior" (early morning) to "mostly data" (late afternoon) without ad hoc parameter choices
• Robust to outlier intervals when the prior variance is appropriately calibrated

Weaknesses

• Computationally more involved than ratio or smoothing methods (though the conjugate Normal case is a few lines of code)
• Requires specifying the prior variance from historical forecast error data — garbage in, garbage out
• Not available as a built-in feature in most commercial WFM platforms as of 2026
• Harder to explain to operations stakeholders than "we multiplied by 1.12"

When to use

Bayesian reforecasting is warranted when the organization needs probabilistic remaining-day forecasts (not just a revised point estimate), when forecast uncertainty is operationally meaningful (e.g., deciding whether to activate overtime), or when the WFM team has the analytical capability to implement and maintain it. Organizations running custom forecasting pipelines in R or Python can implement this with minimal effort using packages like PyMC or Stan.

Method comparison

Criterion	Simple Ratio	Exponential Smoothing	Bayesian Posterior
Implementation complexity	Trivial	Low	Medium
Outlier robustness	Poor — one bad interval distorts everything	Good — smoothing dampens spikes	Good — prior anchors estimate
Output type	Point estimate	Point estimate	Distribution (point + uncertainty)
Tuning parameters	None	Alpha (smoothing weight)	Prior variance (from historical data)
Platform availability	Universal	Common	Rare — typically custom-built
Best scenario	Uniform level shift	Volatile day, noisy early data	Need probabilistic remaining-day forecast

Worked example: 10am reforecast

A contact center forecasted 4,800 calls for the day. The center operates 8am–8pm (48 fifteen-minute intervals). At 10am, 8 intervals have completed.

Original forecast vs. actuals (8am–10am)

Interval	Forecast	Actual	Ratio
8:00–8:15	85	92	1.082
8:15–8:30	110	121	1.100
8:30–8:45	135	149	1.104
8:45–9:00	150	163	1.087
9:00–9:15	160	178	1.113
9:15–9:30	155	172	1.110
9:30–9:45	145	159	1.097
9:45–10:00	140	152	1.086
Total	1,080	1,186	1.098

The morning ran 9.8% hot. Now apply each method.

Simple ratio

Ratio = 1,186 / 1,080 = 1.098. Remaining forecast (original: 3,720) becomes 3,720 × 1.098 = 4,085. Revised daily total: 1,186 + 4,085 = 5,271 (up from 4,800).

Exponential smoothing (alpha = 0.2)

Computing the smoothed ratio interval by interval:

• After 8:00: r = 0.2(1.082) + 0.8(1.000) = 1.016
• After 8:15: r = 0.2(1.100) + 0.8(1.016) = 1.033
• After 8:30: r = 0.2(1.104) + 0.8(1.033) = 1.047
• After 8:45: r = 0.2(1.087) + 0.8(1.047) = 1.055
• After 9:00: r = 0.2(1.113) + 0.8(1.055) = 1.067
• After 9:15: r = 0.2(1.110) + 0.8(1.067) = 1.076
• After 9:30: r = 0.2(1.097) + 0.8(1.076) = 1.080
• After 9:45: r = 0.2(1.086) + 0.8(1.080) = 1.081

Remaining forecast: 3,720 × 1.081 = 4,021. Revised daily total: 1,186 + 4,021 = 5,207.

The smoothed estimate is slightly more conservative than the simple ratio (5,207 vs. 5,271) because exponential smoothing takes several intervals to fully absorb the deviation — the smoothed ratio (1.081) has not yet caught up to the cumulative ratio (1.098).

Bayesian posterior

Prior: daily total $N(4800,300^2)$ where 300 is the historical daily forecast standard deviation. The first 8 intervals represent 22.5% of expected daily volume (1,080 / 4,800). The actual-implied daily total from morning data: 1,186 / 0.225 = 5,271.

Posterior mean (simplified Normal-Normal):

${\mu }_{\text{post}}=\frac{300^2\times 5271+{\sigma }_L^2\times 4800}{300^2+{\sigma }_L^2}$

With observation variance scaled for 8 intervals of data, the posterior mean falls between 5,050 and 5,200 depending on the assumed observation noise — closer to the data-implied estimate than the prior, but pulled slightly toward the original forecast. Critically, the posterior also provides a credible interval: the revised forecast is not "5,150" but "5,150 with 80% CI [4,920, 5,380]."

Automation in WFM platforms

Most commercial WFM platforms provide some form of automated intraday reforecasting:

• NICE IEX / CXone WFM — Intraday Manager computes interval-level reforecasts using a variant of ratio adjustment with configurable look-back windows. Automatic reforecasts can trigger at preset intervals (every 30 minutes, every hour) or on threshold deviation.
• Verint WFM — Provides "Intraday Reforecast" functionality that recomputes remaining-day forecasts using completed actuals. Supports both automatic schedule recalculation and manual override.
• Calabrio WFM — Intraday reforecast engine with adjustable sensitivity. Can be configured to trigger reforecasts when cumulative variance exceeds a threshold (e.g., ±5% deviation from forecast).
• Genesys Cloud WFM — Real-time forecast updates with configurable reforecast frequency and automatic schedule optimization triggers.

The platform implementation matters less than the governance around it: when does a reforecast trigger downstream action? A reforecast that updates the forecast number but does not trigger schedule adjustments is informational, not operational.

Impact on downstream staffing

A reforecast is only valuable if it changes staffing decisions. The chain:

1. Reforecast fires — revised interval-level volume and/or AHT for remaining hours
2. Requirements recalculate — Erlang C (or the site's staffing model) recomputes required positions per interval using revised demand
3. Gap analysis — compare new requirements to current schedule; identify intervals that are now over- or under-staffed
4. Adjustment actions — voluntary overtime offers, early release, break/lunch movement, skill reassignment, callback postponement

The key decision is the threshold: how much reforecast deviation triggers action? Too sensitive (reforecast on every 2% deviation) and the operation spends more time adjusting than producing. Too insensitive (only reforecast on 15%+ deviation) and the operation misses actionable information.

A practical default: reforecast every 30–60 minutes; trigger staffing adjustments when remaining-day forecast changes by more than 5% from the last action point. This balances information exploitation against operational disruption.

Common pitfalls

• Reforecasting too early. A reforecast at 8:30am based on two intervals of data is noise, not signal. Most practitioners wait until at least 2–3 hours of actuals are available before placing confidence in the reforecast.
• Ignoring the cause. A hot morning driven by a system outage (queued calls clearing) will not persist into the afternoon. Blind ratio adjustment will over-staff the afternoon. The forecaster should investigate the why before accepting the reforecast at face value.
• Not reforecasting AHT. Volume reforecasting gets attention; AHT reforecasting often does not. But a day with 10% higher volume and 15% higher AHT requires very different staffing than one with 10% higher volume and normal AHT. Both inputs should be reforecast.
• Reforecast without action authority. If the reforecast cannot trigger staffing changes (because schedule changes require supervisor approval that takes hours), the reforecast is academic. The organizational process must be as fast as the analytical update.

Relationship to other pages

• Forecasting Methods — parent page covering the full method taxonomy
• Forecast Accuracy Metrics — reforecast accuracy should be measured separately from original-forecast accuracy to evaluate the value added by the update
• Real-Time Operations — the operational response that consumes the reforecast output
• Exponential Smoothing — the mathematical foundation for Method 2
• Probabilistic Forecasting — the Bayesian method (Method 3) produces probabilistic output by design