All Articles
Predictive IntelligenceMachine LearningAttrition

How Siwaan Predicts Attrition 6 Months in Advance

A deep dive into the multi-signal XGBoost model, SHAP explainability, and the 40+ data points that power Siwaan's 83%-accurate attrition predictions — and why explainability matters as much as accuracy.

AP

Arjun Patel

AI Research Lead

9 min read
May 28, 2026

Predicting whether a specific person will leave your organization in the next six months is a genuinely hard problem. It's not hard because the data doesn't exist — it's hard because the relevant signals are scattered across a dozen systems, they interact with each other in non-linear ways, and what predicts attrition in one organization may mean nothing in another.

Siwaan's attrition prediction model was designed with this complexity in mind. Here's how it actually works.

Why XGBoost, Not a Neural Network

The first design decision was the choice of algorithm. Neural networks get a lot of attention in AI, but they're poorly suited for structured tabular HR data — the kind that comes from HRIS systems, project tools, and engagement surveys. XGBoost (Extreme Gradient Boosting) consistently outperforms deep learning on this type of data, and critically, it supports explainability in a way that neural networks don't.

Explainability isn't a nice-to-have in workforce intelligence. When a people leader sees a high-risk score on a specific engineer, the first question is always: why? If the answer is 'the model says so,' the insight is nearly useless. The intervention requires understanding the cause. XGBoost + SHAP (SHapley Additive exPlanations) lets us answer that question precisely.

The 40+ Signal Architecture

The model ingests signals from multiple data categories, each connected via Siwaan's zero-code HRIS integrations:

  • Project & utilization signals: bench time duration, utilization rate trend, project assignment frequency, role-to-skill match quality on recent projects
  • Career progression signals: time since last promotion, promotion velocity vs. peer cohort, compensation percentile relative to market and internal band, lateral movement history
  • Engagement signals: survey sentiment trend (not just latest score, but directional change), 1:1 frequency and recency, manager relationship stability
  • Social graph signals: peer departures in the last 90 days, team composition changes, manager changes
  • Learning & growth signals: IDP progress, skill gap trajectory, learning resource engagement, mentor match quality
  • Tenure and lifecycle signals: typical flight-risk windows for the role type, time since last major project milestone

Org-Specific Model Training

One of the most important architectural decisions in Siwaan is that the model trains on each organization's own data, not a generic industry dataset. This matters because what predicts attrition at a 400-person IT consulting firm in Bangalore is different from what predicts it at a 2,000-person cloud services company in Austin.

Within 72 hours of connecting an HRIS, Siwaan has enough historical signal to begin calibrating an org-specific model. The model continues to improve as more data accumulates — particularly around confirmed attrition events that allow for retrospective signal analysis.

83%
Prediction accuracy
Validated on holdout data across real customer orgs, 6-month prediction horizon
40+
Input signals
Across HR, project, engagement, learning, and social graph data
72h
Time to first report
After HRIS connection, first attrition risk scores are available

SHAP Explainability: From Score to Intervention

A risk score of 87% means little without context. SHAP values decompose each individual prediction into the contribution of each input signal, expressed as directional impact on the final score. This means every high-risk profile in Siwaan shows you exactly which signals are driving the score — and by how much.

For example: an engineer with an 87% risk score might show SHAP contributions of: bench time (+34%), sentiment drop (+28%), tenure mismatch (+20%), peer departures (+16%). The people leader can now see that the primary driver is bench time — which has a clear, immediate intervention: project assignment. The secondary driver, sentiment drop, points toward a 1:1 conversation.

What 83% Accuracy Actually Means

Prediction accuracy in attrition models requires careful interpretation. 83% accuracy on a 6-month horizon means that in a validation set of real customers, 83% of eventual departures were flagged as high-risk at least 6 months before the resignation. The false positive rate is kept low by model calibration — generating too many false positives would overwhelm people leaders with unnecessary interventions.

In practice, for a 500-person IT services firm with 10-15% annual attrition, this means catching 8-12 of 10-15 eventual departures 6 months in advance — enough lead time to intervene, have meaningful career conversations, address compensation gaps, or accelerate project placement.

Continuous Model Improvement

The model doesn't stop learning after the initial training period. Each confirmed departure (and each intervention that successfully retained a high-risk employee) is fed back into the training loop. Over 3-6 months of operation, Siwaan's model for a specific organization becomes increasingly calibrated to that org's unique cultural patterns, career trajectory norms, and departure triggers.

See these ideas on your data

Experience the power of prescriptive people intelligence. Schedule a 45-minute live demo using your real HRIS indicators. First attrition risk forecast is included.

Book a Guided Demo