Forecasting Fleet Utilization & Maintenance: A DS View (Redacted)

Note on the data: The figures shown here have been redacted and transformed to protect proprietary aircraft records. Tail numbers identifying specific aircraft have been removed. Original maintenance and flight-hour data were scaled, jittered, and shifted to preserve overall patterns, but the exact dollar amounts and usage figures are not real. Think of it as a “blur filter” for time-series: you can still see the structure of peaks, valleys, and relationships, but the sensitive business details stay private.

Why I did this and why now

I began this analysis in the spring of this year while taking a graduate course on time series analysis, and I have continued to produce monthly reports. This aircraft data was the most obvious time series on my desk at work: two airframes, monthly flight hours, and monthly maintenance costs. The question practically asked itself.

There’s an old line about business aviation: if the jet isn’t flying, it’s hemorrhaging money. That does not mean grounded time causes more maintenance. It means fixed costs are incurred either way, so flying helps offset them, and honestly, some issues are easier to catch in the air than on the tarmac. The practical version of that idea is what I wanted to quantify. If hours are the driver, can I convert them into a month-ahead maintenance schedule that a logistics specialist can actually use to inform budgeting and staffing decisions?

This post guides data analysts and operations teams through aircraft maintenance forecasting, highlighting insights to improve short-term planning and scheduling.

Aircraft and windows Large-cabin Business Jet: April 2021 to July 2025 Small-cabin Business Jet: May 2022 to July 2025

What the data say at a glance

The first pass was decomposition. What remains if trend and seasonality are peeled away from the noise? Then I checked autocorrelation and basic spectra. There isn’t much forecastable seasonality that survives those checks. ARIMA doesn’t capture a strong seasonal pattern either, which already tells you how far you can push a forecast before you leave the lane.

Hours are the driver. Maintenance responds with slight lags and cost pulses. Both lose memory fast. On the large-cabin jet, you can see a quarterly pulse in maintenance, while hours are maintained in short runs. On the small-cabin jet, you see a steady baseline after logging, followed by occasional spikes that are clearly event-driven. Recognizing these differences helps you plan more effectively for each aircraft type.

Short horizons carry a signal because the system has a short memory. Hours cluster for a few months at a time. Maintenance reacts in the same window. That gives you usable guidance for the next one to four months, but beyond that, the patterns become erratic. The large-cabin jet’s quarterly wave drifts in phase. The small cabin never locks into a repeatable rhythm. Real life barges in: vendor slots, parts delays, inspections that slide, and one-off events. At that point, you are not forecasting a pattern so much as trying to time future surprises. Anything beyond four months behaves more like a one-off event than a time-series component.

For all results below, I report 95 percent intervals for horizons 1 through 4. That is the operational window that matters.

The Large-cabin Business Jet story

The large-cabin jet’s quarterly cycle offers a clear planning rhythm, enabling logistics teams to stage parts and schedule work effectively within a two- to three-month window.

Practical translation: stage parts near the quarterly crests, shift lighter work into troughs, and set short labor bands to the 2–3 month rhythm.

What the charts show: the seasonal bump repeats roughly every quarter, the hours ACF stays positive for a couple of lags and then falls off, and the hours→maintenance cross‑correlation peaks at 0–1 months. That is why the same‑month VARX nowcast works and why the next 1–4 months keep their shape before phase drift takes over.

The Small-cabin Business Jet story

The small-cabin jet realization is steady and stationary until it isn’t. After logging, maintenance sits on a routine baseline and then throws spikes that are clearly event‑driven. Hours are choppy, with little persistence and no durable seasonal pattern, so most usable guidance is limited to the short run.

Practical translation: keep tight parts buffers, shorter staffing bands, and plan tactically for one‑off work. Use the model for calls of 1–3 months and avoid over-promising anything beyond that.

What the charts show: decomposition yields a flat trend line with small seasonal fingerprints that do not survive diagnostics. As ACF hours fade quickly, maintenance residuals appear close to white, and the hours→maintenance cross-correlation peaks at 0–1 months and then declines. Same‑month VARX adds a bit when $H_t$ (the realized flight hours in the current month) is known; for $h>1$ the models lean toward baseline behavior fast, intervals widen, and any “cycle” you think you see is wishful thinking. This is why the small-cabin jet remains a 1–3-month tactical instrument, and beyond that, it becomes event-driven.

Methods in plain words

When it came to methods, I kept the philosophy simple: add complexity only if the data warrants it. That makes the workflow more straightforward to follow and the results easier to explain.

The first step is decomposition, which allows us to see how much of the shape is due to trend versus seasonality before any model attempts to fit itself. If the time series appears stable enough on its own (without other drivers) under diagnostics, I turn to ARIMA, selecting orders based on AICc/BIC and stationarity checks. When the real question is how hours drive maintenance, I bring in VAR or VARX: maintenance remains endogenous, with hours as the driver.

That naturally splits the work into two scenarios:

1) Same‑month nowcast: If I already know the current month’s hours, I can plug them straight in. VARX with $H_t$ and a couple of lags yields the same-month maintenance estimate before the invoices are fully received.

2) Month‑ahead forecast: If I don’t know the hours yet, I either run a VAR without them or supply an hours path (average, predicted, etc., there are many ways to do this). This allows the two series to still move together rather than drifting apart.

The usual guardrails apply: checking residuals for whiteness, unpacking the ACF/PACF, and confirming AR and VAR roots are inside the unit circle. The only difference is when the diagnostics say it’s necessary. The simple model wins.

Validation

To validate the analysis, I performed a rolling origin backtest over a six-month window. Both aircraft confirm the short‑horizon picture: the large-cabin jet’s sweet spot is h = 1–4, while the small-cabin jet holds value mainly at h = 1–3 before fading toward noise.

Considerations:

• Forecast intervals provide a sense of the uncertainty surrounding the prediction. A 95% interval means that if we repeated this process many times, we expect the intervals to contain the actual outcome 95% of the time.

• Robustness is more important than a single good-looking fit. If minor tweaks in lag length or differencing change the outcome entirely, the model is sensitive, and this forecast is fixed to the model’s assumptions and data.

• In technical shorthand: autocorrelation for hours and maintenance drops near zero around lags 3 to 4. Cross‑correlation plots only show short leads from hours to maintenance. Backtests converge toward the baselines after a quarter as interval widths approach the unconditional variance. That marks the short‑memory, event‑driven boundary.

What this changes for planning

The results are intended for near-term budgeting and scheduling. They could also provide detailed operational tasks, such as parts ordering or shop assignments, but those fall outside the scope of this analysis. The short‑memory structure gives guidance on how to frame expectations rather than dictating exact actions.

• Budgets: set cost ranges that reflect the short‑horizon signal and avoid false precision beyond a quarter.
• Staffing: adjust labor expectations to the 2–3 month rhythm where it shows up (large-cabin) and keep them shorter and tactical where it does not (small-cabin).
• Scheduling: align higher‑load periods with the large-cabin quarterly pulse and leave contingency room for the small-cabin jet’s one‑off spikes.

The models do not replace operational expertise. However, they provide a consistent way to describe near-term expectations that complements how these aircraft already behave.

Risks, limits, and redaction

Regime shifts and vendor changes rewrite the script. Reporting lag can smear month boundaries. Redaction hides magnitudes so that small effects may look even smaller on the page. The small-cabin jet’s early span is shorter than the large-cabin’s, so any longer cycle would require more history to confirm. Looking ahead, a natural extension would be to bring Chartered Hours into the analysis. That addition could sharpen scheduling forecasts by linking office aircraft use with charter operations. That is the lay of the land, not an apology for modeling.

Want the friendly read or the code

If you are an executive and made it this far, bravo and congratulations. Your proverbial work shoes are probably stained by the proverbial weeds you have been in for the last 30 minutes. That kind of endurance deserves a high‑level version (HTML below). For readers interested in the technical side, the full source is also available.

Open the executive HTML report: 👉 View the report

Browse the source and the rendered artifact: 👉 https://github.com/schwill2018/aircraft-ts-analysis

Implementation notes for the curious: Models used here include ARIMA for univariate structure checks and VAR/VARX for the joint hours-maintenance relationship. VARX was applied for same-month maintenance when hours were known, while for $h > 1$ an hours path was supplied to keep maintenance and hours forecasts aligned. Intervals were reported at the 95% confidence level for horizons 1 to 4, which are the decision windows emphasized in this analysis.