Time Series Analysis

LOWESS for trend extraction and temporal smoothing.

Overview

Time series data often contains noise, seasonality, and trends. LOWESS provides flexible trend extraction without parametric assumptions.


Basic Trend Extraction

fraction = 0.1 sizes the neighbourhood as 10% of the data at each evaluation point — narrow enough to follow a slowly varying trend without smearing periodic variation. Three robustness iterations down-weight noise spikes so they cannot bias the fitted curve; this is especially important when the signal-to-noise ratio is low or when occasional outliers are expected.

using FastLOWESS

n = 500
t = collect(range(0, 100, length=n))
y = [10.0 + 0.5 * t[i] + 3.0 * sin(t[i] / 10.0) + (mod((i - 1) * 7 + 3, 1.7) - 0.85) * 3.0 for i in 1:n]

# Extract trend
model = Lowess(; fraction=0.1, iterations=3)
result = fit(model, t, y)

println("y[0]: ", result.y[1])
y[0]: 11.3116026966887

Detrending

Remove trend to analyze residual patterns.

Setting outputs = ["residuals"] stores observed − smoothed alongside the smooth. A slightly wider fraction = 0.3 produces a smoother baseline trend, so short-duration oscillations end up in the residuals rather than being absorbed into the trend component. The residual series is then ready for spectral analysis, seasonality detection, or change-point methods.

using FastLOWESS

n = 100
t = collect(range(0, 2π, length=n))
y = sin.(t) .+ 0.1

# Smooth to get trend and residuals
model = Lowess(; fraction=0.3, iterations=3, outputs=["residuals"])
result = fit(model, t, y)

trend = result.y
detrended = result.residuals

println("residuals[0]: ", detrended[1])
residuals[0]: -0.15823602487822588

Forecasting with Prediction Intervals

Prediction intervals widen the uncertainty band to include both the uncertainty in the fitted curve (confidence interval) and the expected scatter of new observations around it. fraction = 0.2 offers a balance between local detail and stable interval width — too small a fraction produces jagged interval edges; too large a fraction underestimates local variance near turning points.

using FastLOWESS

n = 100
t = collect(range(0, 2π, length=n))
y = sin.(t) .+ 0.1

model = Lowess(;
    fraction=0.2,
    iterations=3,
    intervals=(confidence=0.95, prediction=0.95),
)
result = fit(model, t, y)

# Intervals are available in result.prediction_lower/upper
println("95% PI: [$(result.prediction_lower[1]), $(result.prediction_upper[1])]")
95% PI: [0.15614437877449766, 0.2989044886023988]

Handling Missing Data

LOWESS naturally handles irregular time sampling:

using FastLOWESS

t_irregular = [i * 1.0 + mod(i * 31, 10) * 0.1 for i in 0:99]
y_irregular = [10.0 + 0.3 * t + 2.0 * sin(t * 0.1) for t in t_irregular]

# LOWESS handles this seamlessly
model = Lowess(; fraction=0.2)
result = fit(model, t_irregular, y_irregular)
println("y[0]: ", result.y[1])
y[0]: 11.32705234623955

Multi-Scale Analysis

Use different fractions to extract features at different scales:

using FastLOWESS

n = 100
t = collect(range(0, 2π, length=n))
y = sin.(t) .+ 0.1

fractions = [0.05, 0.2, 0.5]

results = map(fractions) do f
    model = Lowess(; fraction=f)
    result = fit(model, t, y)
end
for (f, res) in zip(fractions, results)
    println("fraction=$(f): y[0] = $(round(res.y[1]; digits=4))")
end
fraction=0.05: y[0] = 0.1317
fraction=0.2: y[0] = 0.2275
fraction=0.5: y[0] = 0.3325

Gene Expression Time Course

Biological application:

using FastLOWESS

hours = collect(0.0:0.5:24.0)[1:49]
expression = [100.0 * (1.0 + 0.5 * sin(hours[i] * pi / 12.0)) + (mod((i - 1) * 7 + 3, 1.7) - 0.85) * 10.0 for i in 1:49]

model = Lowess(;
    fraction=0.3,
    iterations=3,
    outputs=["diagnostics"],
    intervals=(confidence=0.95,),
)
result = fit(model, hours, expression)

println("R2: ", result.diagnostics.r_squared)
R2: 0.9756292937997523

Choosing Fraction for Time Series

Data TypeRecommended FractionRationale
Daily data (years)0.3–0.5Capture annual trends
Hourly data (days)0.1–0.2Capture daily patterns
Sensor data (minutes)0.05–0.1Preserve short-term features
Noisy dataHigherReduce noise impact
Clean dataLowerPreserve detail

See Also