Imagine opening a dashboard and seeing a list of your top cloud architects, each flagged with an attrition risk score: '87% risk,' '72% risk,' '64% risk.' What do you do next? If you are like most managers, the answer is: nothing. You don't trust the number because you don't know how it was calculated. You can't justify spending retention budget on a hunch, and you can't have a career conversation based on an algorithm's output. The score is interesting, but it is not actionable.
This is the primary failure mode of first-generation HR predictive analytics. Organizations focus on model accuracy while treating the model itself as a black box. But in workforce intelligence, explainability matters more than accuracy. A slightly less accurate model that tells you exactly *why* a person is at risk is infinitely more valuable than a highly accurate model that gives you only a raw score.
Siwaan's attrition prediction engine solves this design gap. By combining a 47-feature XGBoost model with SHAP (SHapley Additive exPlanations) values, we decompose every individual prediction into a ranked list of specific drivers, turning black-box scores into plain-language narratives that managers can trust and act on.
Why Black-Box Scores Create Distrust and Inaction
When predictive models are implemented as black boxes, they fail to drive change. Managers ignore the flags because they lack reason codes. HR Business Partners (HRBPs) struggle to coordinate interventions without knowing whether the driver is compensation, workload, or career stagnation. This lack of transparency leads to three predictable outcomes:
- Inaction. Since managers cannot explain the prediction, they wait for the employee to resign before intervening -- which is too late and too expensive.
- Mistargeted spend. Organizations apply generic retention measures (e.g., broad compensation adjustments) that waste budget on employees who are leaving for non-financial reasons.
- Erosion of trust. A few false positives are enough for managers to dismiss the predictive system entirely, returning to gut-feel decision-making.
The 47 Features Across Six Signal Domains
Siwaan's model integrates 47 features spanning six key workforce domains to capture the complex, non-linear signals that precede attrition:
1. Project and Utilization Signals
Bench duration, utilization rate trend, project assignment frequency, role-to-skill match quality, context switch count, involuntary bench flag, project complexity score, and allocation stability.
2. Career Progression Signals
Time since last promotion, promotion velocity vs. peer cohort, compensation percentile, lateral movement history, title-responsibility alignment, career path clarity score, IDP completion rate, goal achievement trend, and skill ceiling proximity.
3. Engagement and Sentiment Signals
WRS sentiment trajectory, survey score trend, 1:1 meeting frequency, manager relationship stability, blocker mention rate, discretionary effort signals, and communication pattern changes.
4. Social Graph and Team Dynamics
Peer departures in the last 90 days, team composition changes, manager changes, reporting line stability, mentoring relationship presence, and collaboration network density.
5. Learning and Growth Signals
GVI trajectory, skill gap velocity, learning resource engagement, mentor match quality, certification progress, cross-functional project exposure, stretch assignment history, and technical depth-vs-breadth balance.
6. Compensation and Market Signals
Comp ratio to market, comp ratio to internal band, variable pay trajectory, benefits utilization, external market demand for skill set, competitor hiring velocity in region, industry attrition benchmark, geographic cost-of-living delta, and last comp adjustment recency.
Why XGBoost, Not Neural Networks
We chose a gradient-boosted decision tree algorithm (XGBoost) over deep neural networks for two primary reasons. First, tree-based models perform significantly better on structured, tabular HR data, which contains missing values and categorical features. Second, XGBoost is natively compatible with SHAP feature attribution techniques, allowing us to compute exact contribution scores for each feature on every individual prediction.
SHAP Explainability -- From Score to Narrative
SHAP values allow us to decompose a risk score (e.g., 87% risk) into the precise contribution of each feature, measured relative to the organization's baseline risk. The platform translates these mathematical values into clear, written narratives:
- Narrative A: 'Risk score is 87%, driven by 45 consecutive days on the bench (+34%), declining WRS sentiment (+28%), and time since last promotion (+20%).'
- Narrative B: 'Risk score is 72%, driven by a compensation ratio in the 30th percentile of market (+41%), peer departures on their project team (+21%), and manager change in the last 90 days (+12%).'
- Narrative C: 'Risk score is 64%, driven by high context-switching between three projects (+30%), sustained utilization over 95% for three months (+24%), and a drop in 1:1 meeting frequency (+10%).'
The Intervention Playbook
By linking SHAP reason codes directly to specific interventions, Siwaan automates the creation of retention playbooks. If Narrative A's top driver is bench time, the playbook recommends project staffing. If Narrative B's top driver is compensation, it prompts a comp adjustment review. If Narrative C's top driver is context-switching, it triggers an allocation rebalancing. This transforms retention from a reactive guessing game into a targeted, data-backed discipline.
The ROI of Proactive vs. Reactive Retention
Proactive retention is significantly cheaper than reactive retention. When an employee hands in their resignation letter, their manager is forced to make a counter-offer under pressure. Counter-offers are expensive, create internal compensation disparity, and have a low long-term success rate -- over 60% of employees who accept a counter-offer leave within 12 months anyway.
By predicting flight risk six months in advance, Siwaan gives organizations time to address the root causes of disengagement -- career stagnation, project mismatch, or manager friction -- long before the employee decides to interview elsewhere.
We used to get risk scores and ignore them because we didn't know what to do. With Siwaan's SHAP breakdowns, I know immediately if someone is leaving because they are bored on the bench or because they feel underpaid. I can have an honest conversation and make a targeted change that actually works.
— Senior HRBP, Professional Services Organization
The Siwaan Approach
Siwaan does not offer black-box AI. Our attrition engine is built on absolute transparency. Every prediction includes a complete SHAP-based breakdown of inputs, mapped directly to actionable playbooks. The model trains on your own data and runs locally inside your secure tenant boundary, ensuring that your talent strategies remain private, fair, and highly effective.