060 279 5587 info@sitect.co.za 139 Davies Street, Doornfontein, Johannesburg, 2001 Gauteng, SA
Services / AI Consultation & Integration / Predictive Analytics & Forecasting
Custom ML · Confidence intervals · Live dashboards

Stop guessing. Start forecasting. With confidence intervals.

Custom machine-learning models trained on your data — revenue, demand, churn, anomalies — delivered as live dashboards and APIs your team can act on. Every forecast comes with a confidence band, a backtest report, and drift monitoring so you'll know before the model starts to slip.

Live in 6 weeks Backtested + reproducible Drift-monitored From R22,000

Four ways teams forecast — and end up flying blind.

If your "forecast" is a spreadsheet someone updates on a Sunday, you'll recognise these. We've cleaned up plenty.

!

The Excel oracle

A junior analyst extrapolates last year's growth, hard-codes it as a percentage, and the board treats it as a fact. No seasonality. No drivers. No confidence.

=B2*(1+15%)
!

One number, no range

"We'll do R12.4M next quarter." Will you? With what probability? Plus-or-minus what? A forecast without a confidence interval is a fortune cookie.

"Q3 revenue = R 12,400,000"
!

Lagging only

Reports tell you what already happened. Forecasts tell you what's coming. Most BI tools do (a) brilliantly and (b) not at all. You need both.

"YTD revenue: R 38M ✓"
!

Black-box, no audit

An AutoML tool spits out a prediction. Nobody can say why it predicted that. When it's wrong (and it will be), nobody knows what to fix.

// model_v17.pkl · trained: ???

Eight forecast types. All custom-trained.

If the question is "what will happen?" we can probably build a model for it. These eight cover ~85% of what SA businesses ask for.

Revenue forecast

Daily, weekly, monthly revenue by product/region/channel. With seasonality + promo lifts.

Horizon30 – 365 days

Customer churn

Per-customer probability of churning in the next 30/60/90 days. With top features driving the risk.

Horizon30 – 90 days

Demand forecast

SKU-level demand for inventory + production planning. Right size your stock without stockouts.

Horizon1 – 12 weeks

Lead scoring

Per-lead probability of converting + estimated deal size. Route the hot ones to your best reps.

HorizonReal-time

Anomaly detection

Catch fraud, ops outliers, data quality issues before they hit the bottom line. Streaming + batch.

HorizonReal-time alerts

Capacity / staffing

Call-centre volume by half-hour, store footfall, ops team load — staff to demand instead of guessing.

HorizonHourly – 12 weeks

Price optimisation

Elasticity models that recommend the price-per-product to maximise revenue, margin, or volume.

HorizonContinuous

Cash-flow forecast

Daily cash position + 13-week rolling forecast. Knows your debtors' DSO patterns and seasonality.

Horizon13-week rolling
What you actually get

Predictions, but also why.

Every Sitect model is auditable. You see the prediction, the confidence interval, the top features driving it — and a backtest of how accurate the model would have been on every past month.

  • Live dashboards — embed in Looker / Metabase / your custom app via signed URLs.
  • Confidence bands — 80% and 95% intervals as standard, configurable.
  • Feature attributions — SHAP values per prediction so you know what's driving the number.
  • Drift detection — alerts the day the model's input distribution shifts beyond a threshold.
  • Backtest report — see how well the model would've done on the last 24 months, before it goes live.
  • API endpoint — call /forecast from your app and embed predictions wherever they matter.

The right tool for each forecast.

We don't have a religion about modelling. We pick from the methods below based on your data shape, the horizon, and the explainability you need.

Classical time-series
Strong baseline · explainable

Solid baseline for trends with clear seasonality (weekly, monthly, holiday patterns). Fast to train, easy to explain to non-technical stakeholders.

Prophet ARIMA SARIMAX ETS statsforecast
Gradient-boosted trees
Tabular · feature-rich

Workhorse for forecasts with lots of features (price, weather, promos, day-of-week). SHAP values give per-prediction explainability your CFO can challenge.

XGBoost LightGBM CatBoost SHAP
Neural forecasts
Complex patterns · multivariate

For high-volume, long-horizon, or multivariate problems where classical models leave accuracy on the table. Used selectively — we don't over-engineer.

NHITS TimesNet PatchTST N-BEATS
Classification & scoring
Churn · leads · risk

When the question is "how likely is X to happen for this user?" we use probabilistic classifiers calibrated to deliver real probabilities — not just rankings.

Logistic Reg XGBoost-Class Calibration
Anomaly detection
Fraud · ops outliers

Unsupervised detection of unusual patterns — transactions, sensor readings, conversion drops. Configurable per-metric thresholds, streaming or batch.

Isolation Forest LOF Z-score CUSUM
Ensembles + LLMs
Blended + narrative

Best predictions usually come from blending. We weight models by their backtest performance, and use an LLM to translate the numbers into board-ready commentary.

Stacking Weighted avg GPT-4o narrative

Data in. Decision out.

Five stages from raw data to your dashboard. Every stage is versioned in Git, every model run is reproducible, every prediction is logged.

01 · CONNECT

Data sources

Read your existing systems.
  • Xero · SAP · Sage
  • Postgres · MySQL · Snowflake
  • Salesforce · HubSpot
  • S3 · GCS · CSV
02 · FEATURES

Engineer

dbt + SQL · versioned.
  • Lags · rolling means
  • Seasonal flags
  • Promo lifts
  • Customer cohorts
03 · TRAIN

Backtest + train

Reproducible runs.
  • Cross-validation
  • Hyper-param search
  • MLflow registry
  • Backtest report
04 · SERVE

Predict

API + scheduled jobs.
  • REST API
  • Scheduled refresh
  • Streaming option
  • Cached responses
05 · MONITOR

Drift + alerts

Know when to retrain.
  • Input drift PSI
  • Output drift
  • Accuracy alerts
  • Auto-retrain rules
Every model has a backtest report + MLflow run ID · reproducible with one command. Drift detected → Slack alert within 1 hour · auto-retrain available · POPIA-compliant feature store.

Six artefacts. All reproducible.

You don't pay a per-prediction fee or a SaaS subscription. The pipeline runs in your infrastructure; provider tokens billed direct.

01 · Backtest report

Honest accuracy, before go-live

How accurate the model would have been on the last 24 months — MAPE, MAE, RMSE, plus the cases where it would've been most wrong. Read before signing off.

PDF + interactive notebook
02 · Trained model

Versioned + reproducible

The model file + the exact code + the exact data slice that trained it. Re-run in one command. Roll back to any prior version instantly.

MLflow + DVC
03 · Forecast API

REST endpoint your app calls

POST request, JSON response. Returns the prediction, the confidence interval, the top features driving it. Sub-100ms typical latency.

POST /forecast
04 · Dashboards

Pre-built or embedded

Metabase / Looker / custom React component — pick one. Pre-built tiles for the forecast, confidence bands, top features and recent prediction history.

Metabase · Looker · React
05 · Drift monitoring

Slack-alerted retraining

Input data PSI score tracked daily. Output distribution monitored. When either drifts, you get a Slack alert + auto-PR to retrain.

Slack + email + PagerDuty
06 · Runbook + training

Hand-over your team can run

Markdown runbook for the top incidents: model accuracy drop, drift fired, data source down. 2-hour live training for your data + product teams.

Live training + runbook

6 weeks. 5 phases. One trustworthy forecast.

You'll spend ~12 hours of stakeholder time across the whole build. Most of it in week 1 (data audit) and week 5 (backtest review).

01
Week 1 · 4 hrs

Data audit

Map the data. Find the gaps. Quantify the signal. Stop here if the data is too sparse — saves you R22k.

02
Week 2 · 2 hrs

Features

Build the feature pipeline in dbt + SQL. Version it. You can read it and challenge it.

03
Week 3–4 · 2 hrs

Train + backtest

Multiple model families benchmarked. Backtest report drafted. You see the wins and losses before sign-off.

04
Week 5 · 2 hrs

Ship

API deployed, dashboard live, drift monitor armed, Slack channel connected. Pilot users using it.

05
Week 6 · 2 hrs

Tune + handover

Adjust thresholds, train your team, document the runbook. 90 days post-launch monitoring included.

The numbers behind a forecast done right.

Median results across our deployed forecasts. Your numbers will be yours — these are the order-of-magnitude bracket to plan against.

4.2%
Median MAPE
Mean absolute % error on revenue / demand forecasts.
62%
Inventory drop
Median reduction in safety stock without stockouts.
3.7×
Decision speed
From quarterly board pack to weekly auto-refresh.
R890k
Annual saving
Median first-year value across deployed forecasts.

Pick the scope. We deliver in 5–8 weeks.

Build fee is fixed in writing. Compute + cloud hosting bills direct to your AWS / Azure / GCP — typically R800–R3,500/mo depending on data volume. No per-prediction fee.

Single forecast

Starter model

R22,000
/project · ex 15% VAT
  • 1 forecast (e.g. revenue or churn)
  • 1 data source
  • Backtest report
  • REST API + Metabase dashboard
  • Basic drift monitoring
  • 30 days post-launch tuning
Choose Starter →
Enterprise

Forecast factory

R100,000+
scoped per engagement · ex VAT
  • Unlimited forecasts
  • Feature store + model registry
  • Streaming + batch options
  • SSO + audit + RBAC
  • Auto-retrain pipelines
  • SLA + dedicated PM
  • 90 days hyper-care + QBR
Talk to us →

Common questions from CFOs & heads of data.

If yours isn't here, send us a sample data extract — we'll come back with a feasibility note in 48h. No sales pitch.

How much data do we need?
Rough rules: 2+ years of monthly data for revenue forecasts (so the model sees seasonality twice). 500+ customers for churn, with at least 50 known churners. 50+ data points per SKU for demand. For anomaly detection, much less — sometimes 3 months is enough. If we audit your data and it's too thin, we stop the engagement and refund. We've done this 3 times — it's better than shipping a model that lies to you.
What's MAPE / why do you keep mentioning 4.2%?
MAPE = Mean Absolute Percentage Error. The lower the better. 4.2% means the forecast is, on average, within 4.2% of the actual value. For revenue forecasts that's very good — most businesses run on much worse internal forecasts (15–30% MAPE). The exact MAPE you'll achieve depends on your data signal-to-noise ratio; we tell you in the backtest report before you sign off.
Black box or explainable?
Always explainable. Every Sitect forecast ships with SHAP feature attributions per prediction — meaning for any number on the dashboard, you can drill down to "what specifically pushed this number up or down." Your CFO will ask. We make sure your team has the answer.
POPIA — what about customer-level predictions?
Customer-level forecasts (churn, lead score, credit risk) are classified as automated decision-making under POPIA. We build them with: (1) a human-in-the-loop review for high-stakes decisions, (2) the right to explanation built into the dashboard, (3) PII pseudonymisation at the feature layer. We also draft your data protection impact assessment (DPIA) for the engagement.
What if our data has gaps / quality issues?
Common — it's part of the engagement. The audit in week 1 quantifies the gaps. Small gaps (5–15%) we impute robustly. Large gaps mean either a smaller-scope forecast or a recommendation to fix the data pipeline first (we can do that too, as a separate engagement). We won't paper over bad data.
Will the model just keep working, or does it decay?
It decays. Every model has a shelf life — usually 3–12 months before accuracy drops noticeably. The drift monitor catches this automatically and alerts. Retraining is built into the pipeline; for most clients it's a 1-click action after the alert. Some teams set up auto-retrain on a schedule.
Can the model do "what-if" scenarios?
Yes for many model types. Scenario sliders in the dashboard let you change inputs (e.g. price, marketing spend, headcount) and see the predicted impact. Works best with feature-engineered tree models (XGBoost / LightGBM); less clean with deep-learning models. We tell you which approach fits your use case.
What if the forecast is wrong and we make a bad call?
Three layers: (1) Confidence intervals — every forecast is a range, not a single number. (2) Backtest visibility — you see historical accuracy before trusting it. (3) Human override — every dashboard lets your analysts adjust the forecast (with an audit log). Forecasts inform decisions, they don't replace them.

Send us a sample of your data.
We'll send back a feasibility note.

Drop a CSV or a SQL extract into our secure form. Within 48 hours one of our seniors writes a one-pager: signal strength, expected MAPE, recommended approach, indicative timeline + price. No sales pitch — sometimes the answer is "your data isn't ready yet, fix this first."

You'll get back in 48h

  • Data signal-strength score
  • Expected MAPE range
  • Recommended model family
  • Indicative timeline + price

More AI Consultation services

AI Strategy & Roadmap

Where does AI fit in your business?

AI Chatbot Integration

Smart chatbots that capture, qualify and convert.

AI Voice & Phone Agents

24/7 AI agents that handle calls.