Key takeaways
- The Question: What does predictive analytics actually mean on a construction project, what data does it need, which methods work, and how far can you trust a forecast?
- The Answer: It means projecting finish dates and final costs from how fast work is really moving. Simple run-rate and earned value projections cover most needs; statistical models and machine learning add value only with consistent history. Every method stands or falls on objective, frequent measurement of installed work, and a forecast is only useful if someone acts on it.
- The YelloSKYE Angle: Construct IQ supplies the measured side of the forecast: dated, location-tagged progress compared with the 4D schedule, and volumes measured from drone surveys.
Predictive analytics in construction is the use of measured project data, chiefly installed quantities over time, cost to date and the schedule, to estimate where an activity, a milestone or the whole project is likely to finish. Its value is timing. A delay seen four weeks into an eight-week activity can still be recovered; the same delay seen on the planned finish date cannot.
This guide explains what a construction forecast answers, the data behind it, the main methods from simple to advanced, the leading indicators that move first, the limits on accuracy and how to turn a forecast into a decision. The measured inputs are where most forecasts fail, which is why we also show how Construct IQ feeds them from the site.
What a construction forecast answers
A schedule tells you what was intended. A progress report tells you what has happened. A forecast answers a third question: if things carry on as they are, where will we land? On a project that question comes in four sizes.
| Question | Typical horizon | Who uses the answer |
|---|---|---|
| When will this activity in this zone finish? | One to six weeks | Site manager, trade foreman |
| Will we hit the next milestone? | One to three months | Project manager, planner |
| What will the project cost at completion? | Remaining project life | Commercial manager, owner |
| Which projects in the portfolio are drifting? | Quarter to year | Operations director, investors |
The smaller questions are easier to answer well, and they add up. A good milestone forecast is mostly a set of good activity forecasts combined through the logic of the programme.
The data a forecast needs
| Input | What it is | Where it usually comes from | Common weakness |
|---|---|---|---|
| Planned quantities | How much of each activity exists, by zone (square metres of blockwork, metres of duct, number of fixtures) | Bill of quantities, BIM model, take-off | Activities defined too coarsely to measure |
| Measured progress | How much has actually been installed, and when | Site walks, reports, reality capture | Subjective percentages; long gaps between updates |
| Planned dates and logic | Start, finish and dependencies | Scheduling software | Logic missing or out of date |
| Cost to date | Committed and spent cost per package | Cost and accounting systems | Posting lags weeks behind the work |
| Resources | Crews and plant on site | Daily reports, subcontractor returns | Self-reported and inconsistent |
| Constraints | Open queries, late deliveries, blocked access | Query logs, procurement trackers | Kept in separate tools nobody links |
| History | Achieved rates from past projects | Previous project records | Rarely captured in a usable form |
Measured progress is the input that matters most. A forecast built on percentages typed into a spreadsheet inherits every bias in those numbers. Our guide to percent complete in construction covers how to measure progress so that it holds up.
Forecasting methods, from simplest to most advanced
Run-rate projection
The simplest forecast divides the remaining quantity by the recent rate of installation. Take an illustrative blockwork package: 4,800 m² in total, 2,100 m² laid after six weeks, so 2,700 m² remains. Over the last four weeks the crew has averaged 310 m² a week. Seven weeks remain in the plan, so the crew needs about 386 m² a week. At the current rate the work takes almost nine weeks, roughly two weeks late.
That calculation takes a minute, and it is more honest than a reported "45 per cent complete, on track". Its weakness is that it assumes the recent pace continues.
Trend-adjusted projection
Pace changes. Crews speed up as they learn a repetitive floor, then slow down near the end as the awkward areas remain. Weighting recent weeks more heavily, or fitting a curve to cumulative progress, captures this. The classic S-shaped progress curve is a useful reference: if an activity is past its midpoint and the weekly rate is already falling, the finish will usually be later than a straight-line projection suggests.
Earned value forecasts
Earned value management turns progress and cost into indices, and from them into an estimated cost at completion and an implied finish. It is the standard method for whole-project cost forecasting, and it works best when earned value comes from measured quantities. Our earned value management guide sets out the formulas and a worked example.
Schedule risk analysis
Instead of one duration per activity, a schedule risk analysis gives each activity a range (optimistic, most likely, pessimistic) and runs the programme thousands of times in a Monte Carlo simulation. The output is a probability of finishing by each date, often reported as P50 and P80 dates. Its quality depends on whether the ranges reflect real performance or guesses made in a workshop.
Reference class forecasting
Developed from the work of Daniel Kahneman and Amos Tversky and applied to projects by Bent Flyvbjerg, reference class forecasting starts from how a group of similar past projects actually performed, then adjusts for what is different about this one. It corrects the optimism that comes from planning a project from the inside. It needs a reliable record of outcomes from comparable work.
Statistical and machine learning models
With consistent data from many past projects, a model can learn which combinations of factors, such as package type, early pace, query volume and crew size, tend to precede a slip. These models can spot patterns people miss. They also need a lot of clean, comparable history, can be hard to explain to a project team, and fail quietly when a new project does not resemble the training data. Most construction firms do not yet hold that kind of history, so these models remain the exception rather than the rule.
| Method | Data needed | Best horizon | Strength | Weakness |
|---|---|---|---|---|
| Run rate | Quantities and recent progress | Weeks | Fast, transparent | Ignores changes in pace |
| Trend-adjusted | Weekly progress history | Weeks to months | Captures slowing and speeding up | Needs several weeks of data |
| Earned value | Progress plus cost | Whole project | Links time and cost | Sensitive to how progress is credited |
| Schedule risk analysis | Ranges for each activity, sound logic | Milestones | Shows probability, not one date | Ranges are often guessed |
| Reference class | Outcomes of similar projects | Whole project | Corrects optimism | Requires a reliable reference set |
| Machine learning | Large, consistent history | Varies | Finds complex patterns | Hard to explain; needs clean data |
Leading indicators that move before the dates do
Dates slip last. Several measures move earlier and are worth watching every week.
| Indicator | What it tends to predict | How to measure it |
|---|---|---|
| Actual pace against required pace | Activity finish slip | Installed quantity per week compared with remaining quantity divided by remaining weeks |
| Open work areas per trade | Trade stacking and a slow finish | Count of zones each trade has started but not completed |
| Age of open design queries | Stoppages in specific areas | Days each query has been open |
| Weekly plan reliability | Short-term schedule slip | Share of committed weekly tasks completed as promised |
| Materials on site against need dates | Idle crews | Delivery records against the lookahead plan |
| Rework found per inspection | Lost time and cost | Defects raised and closed per zone |
| Zones with no visible change | Stalled work | Comparing successive dated captures of the same area |
None of these need sophisticated software. They need consistent measurement and someone who looks at them on a fixed day.
How accurate can a forecast be?
Forecasts deserve trust in proportion to their track record. Several limits apply to every method.
- Early forecasts are noisy. Two weeks of data on an eight-week activity can swing the projected finish by weeks. Show a range until the trend settles.
- The input sets the ceiling. Subjective progress produces confident forecasts that are wrong. Measured quantities do much better.
- Projects change. A design change, a new subcontractor or a resequence breaks the pattern the forecast was built on. Restart the trend after any structural change.
- Forecasts change behaviour. A good forecast prompts action, which then makes it look wrong. That is the point. Record what was done so the team can see the effect.
- Single dates mislead. "Finish 14 March" implies precision nobody has. "Likely early March, with risk into April" is more honest and more useful.
The simplest discipline is to log each forecast and compare it with what happened. After a few months the team knows how much weight its own forecasts deserve.
Turning a forecast into a decision
A forecast that nobody acts on is just a more detailed report. A short routine helps.
- Confirm on site. Check that the trend reflects reality and not a reporting gap.
- Find the cause. Crew size, missing information, materials, access or sequencing each lead to a different response. Our guide to construction risk management covers how to keep risks tied to evidence.
- Choose a response. Add crews, open a second work front, change the sequence, remove a constraint, or accept the slip and re-plan the work that depends on it.
- Check the effect on downstream work. A recovery plan for one trade can create a clash for the next.
- Set a review date. Decide when you will check whether the required pace is being met, and who owns that check.
Where site capture fits
Every method above depends on knowing how much work is really in place. That is the part YelloSKYE handles. Using drone flights, 360° walks, fixed cameras and laser scans, Construct IQ sets what has been built beside the 4D schedule baseline, so installed work is measured from evidence rather than estimated from memory. For earthworks, drone surveys produce terrain models from which cut and fill are calculated; in our earthwork volumetric analytics use case, a survey cycle that took two weeks came down to a few hours, which means quantity data arrives while it can still shape the forecast.
At portfolio level, DEX shows stage-wise completion and lag across every site in one place. In our portfolio monitoring use case, capturing every site with identical parameters let leadership spot common bottlenecks weeks earlier than manual reporting allowed. The forecasting itself can sit in your scheduling or cost tools; progress data exports to Procore and Autodesk Construction Cloud.
Getting started
- Break critical activities into measurable quantities by zone.
- Measure installed quantities on a fixed weekly cadence, from site evidence where possible.
- Calculate required pace and actual pace for every critical activity each week.
- Track three or four leading indicators alongside the programme.
- Express forecasts as ranges and log them, then compare with outcomes.
- Agree in advance who acts when a forecast crosses a threshold, and how quickly.
- Add schedule risk analysis or models only once the basic measurements are reliable.
