Bachelier and His Legacy

Teams need a simple baseline for uncertain change, but a baseline becomes dangerous when convenience is mistaken for reality. Brownian motion is powerful precisely because its assumptions are clear enough to test.

Enterprise problem: analysts must use a tractable diffusion model without erasing jumps, changing volatility, or dependence that control real failure.

Whole-course context: after survival and dependence, the course now examines the mathematical baseline that shaped modern finance before challenging its limits.

Today’s slice: this chapter reconstructs Bachelier’s random price path, diffusion scaling, and normal increments as a disciplined null model.

End-of-day evidence: you will calculate multi-step variance, operate path and surface labs, and state which observations reject the baseline.

Still unsolved: a Brownian benchmark does not explain heavy tails, volatility clusters, discontinuities, or multifractal trading time.

Diffusion language can sound more mysterious than its mechanics. These terms make the assumptions inspectable.

Term	Plain meaning
Random walk	A path built by accumulating random steps
Brownian motion	A continuous-time random path with independent, stationary Gaussian increments
Increment	Change over a stated time interval
Gaussian	Bell-shaped probability distribution specified by mean and variance
Drift	Average change per unit time
Variance	Expected squared distance from the mean
Diffusion	Spreading of a distribution as time passes
Stationary increments	Increment law depends on interval length, not calendar location
Null model	Simple reference model tested against data

Brownian motion is continuous but nowhere smoothly differentiable in the mathematical ideal. Its rough path was revolutionary, yet its one fixed scaling rule remains milder than many market records.

Financial theory needed a way to price uncertainty before modern probability had settled its vocabulary. In 1900, Louis Bachelier modeled speculative prices as random motion and connected price changes to the mathematics later associated with physical diffusion.

The insight was not that markets are meaningless. If publicly available information is already reflected in price, the next change can be difficult to anticipate. A fair-game baseline prevents analysts from inventing patterns in noise. It also enables probability calculations for contingent claims.

Mandelbrot treats Bachelier with respect because a bold approximation can open a field. The criticism comes later: a successful baseline should invite measurement, not immunity. When observed tails, jumps, or clustering contradict its assumptions, the model must yield.

A Brownian price-change model starts with independent increments whose mean and variance grow with elapsed time. In arithmetic form, an interval change can be written ΔP = μΔt + σ√Δt Z, where Z is standard normal.

The square-root term is the key scaling rule. If one-day variance is σ², the variance over n independent days is nσ², while standard deviation is σ√n. Directional uncertainty cancels in expectation but accumulates in spread.

Modern asset models often apply diffusion to log price so price stays positive: dS/S = μdt + σdW. This geometric Brownian motion underlies Black–Scholes. Bachelier’s arithmetic process can cross zero, which is unrealistic for many assets but remains useful for some price differences and as a historical baseline.

A diffusion distribution spreads smoothly from one starting point. The model links path uncertainty and horizon through one stable volatility parameter.

The model is self-similar in distribution: magnifying time and rescaling amplitude by the square root restores the same statistical form. That is a precise claim. It can be checked using aggregated increments, quantiles, and moments rather than visual resemblance.

Forecast users often multiply daily standard deviation by days instead of the square root of days. A simple calculation shows the difference.

Assume zero drift and daily standard deviation of 2%. Under independent Gaussian increments:

five-day standard deviation = 2% × √5 ≈ 4.47%

It is not 10%. A one-standard-deviation band is roughly -4.47% to +4.47%; a two-standard-deviation teaching band is roughly -8.94% to +8.94%. Those bands depend completely on stable volatility, independent increments, and Gaussian tails.

If daily volatility doubles after day two, variance becomes 2² + 2² + 4² + 4² + 4² = 56 squared-percent units, so standard deviation is √56 ≈ 7.48%. One constant σ would understate the changed regime.

One diffusing path can encourage narrative unless it is compared with a fixed center. The 2D lab isolates one diffusion control against that reference.

Use the lab in controlled comparisons:

1. Press Reset and note the Diffusion percentage.
2. Identify the Fixed mean and Diffusing path curves from their labels and encodings.
3. Increase Diffusion one step and compare separation from the fixed mean at equal time positions.
4. Record the Path spread metric at low, default, and high settings.
5. Reset and confirm that the default curves and spread return exactly.

These are deterministic teaching simulations, not market forecasts. A path ending near the center is no more “correct” than one in a tail; the plotted curves and controlled readout are the objects to interpret.

Diffusion depends jointly on time, volatility, and displacement. A 3D surface shows how the displayed density shape changes across time and displacement without claiming that the rendered finite grid is a normalized probability distribution.

Operate it carefully:

1. Read the legend: x is time, y is displacement, and z is density.
2. Move the Volatility slider and compare the surface range at two settings.
3. Use the x focus slider to choose a time and the y focus slider to choose a displacement; record selected x, y, and z.
4. Hold time fixed while moving across displacement, then select a later time and repeat the slice.
5. Orbit by pointer or keyboard to inspect the shape, then Reset the volatility, focus, and camera.

The rendered mesh samples a formula on a finite grid. Smoothness is a property of the model and interpolation, not proof that observed prices move continuously.

Brownian motion fails as a complete market description when data violate its independent, stationary, Gaussian, continuous increments. Each adjective supplies a separate test.

Independence can be challenged with return and magnitude dependence. Stationarity can fail when volatility changes by hour or regime. Gaussianity can fail through excess tail observations. Continuity can fail through overnight gaps or sudden repricing. Measurement at very short horizons also includes bid–ask bounce and discreteness that no continuous path represents.

Do not discard the baseline merely because it is imperfect. Compare it with richer models and ask which decision changes. A simple model that bounds ordinary behavior can coexist with stress scenarios for what it omits. Model risk begins when a convenient baseline sets limits, capital, or autonomy without those stresses.

Performance teams need a null model for measurement noise before blaming every fluctuation on code. Independent Gaussian jitter can provide that baseline, but production latency often violates it.

If request latency deviations were independent with stable variance, averaging n observations would reduce standard error by 1/√n. Queue buildup, garbage collection, and shared dependencies instead create persistence and heavy right tails. A five-minute average can therefore look stable while users encounter clustered timeouts.

Use the diffusion baseline to calibrate ordinary measurement variation, then test residuals for runs and tail excess. Design timeouts and capacity from observed tail and dependency behavior, not the bell curve alone.

Repeated token sampling creates branching output paths, tempting teams to treat generation as a simple random walk. The analogy helps with accumulation of uncertainty but breaks when state and probabilities change after every token.

In Brownian motion, increments are independent and stationary. In a language model, the next-token distribution depends on the entire prefix, tools, and decoding policy. Small wording changes can redirect later probabilities, and safety failures are not symmetric deviations around a mean.

Use independent sampling as a baseline for repeated evaluations at fixed prompts. For full conversations, preserve conditional structure and report outcome categories, costs, and tails instead of forcing a continuous Gaussian story.

Agent planners explore uncertain paths, but tool outcomes and accumulated state make increments dependent. A Brownian baseline can still expose how rapidly uncertainty would spread under the simplest assumptions.

Suppose independent timing noise adds variance at every tool call. Total timing variance grows roughly with call count, but retries and rate limits can create much faster bursts. State constraints also introduce absorbing boundaries: an irreversible deletion cannot be canceled by an opposite later step.

Use the baseline to budget ordinary variability, then add explicit failure states, correlated tool outages, and policy boundaries. Agent safety depends on discrete consequences that a smooth diffusion surface cannot encode alone.

Startup plans often display a smooth fan of possible revenue paths around one growth drift. That is a diffusion picture, useful for communicating uncertainty but fragile when customer acquisition changes regime.

If weekly growth shocks were independent and stable, forecast variance would accumulate predictably. A platform policy change, major customer loss, or viral loop changes drift and volatility together. The true forecast needs discrete branches for these mechanisms.

Keep the diffusion fan as the base case, add named regime and runway stresses, and update parameters out of sample. A broad fan does not automatically include the right failure mechanism.

People overreact to routine variation and underreact to persistent shifts. A simple diffusion baseline helps ask whether a change is larger or longer than ordinary noise.

Daily weight, commute time, or mood can vary around a local level. Averaging reduces independent measurement noise, but illness, schedule changes, and seasonality shift the process itself. Treating every observation as a trend creates anxiety; treating a sustained shift as noise delays response.

Record context, use rolling summaries cautiously, and define thresholds for gathering more evidence. The model organizes observation; it does not diagnose health or prescribe action.

A Brownian model earns trust only by surviving predictions it did not fit. This exercise converts its four core assumptions into a compact falsification report.

Choose a cleaned series and a base interval. Estimate drift and variance on an initial window, then predict the variance at two, four, and eight intervals using additivity. Compare observed aggregated variance with nσ², and compare standardized quantiles with a Gaussian reference. Check signed and absolute increment dependence separately. Mark any discontinuity that cannot be explained by the sampling protocol.

Repeat the analysis on a later window without changing parameters. Classify each mismatch: independence, stationarity, Gaussianity, or continuity. Do not repair all mismatches at once. Add the smallest mechanism—time-varying variance, a jump component, or a seasonal term—and reserve another statistic for validation.

Create a compact evidence matrix:

Baseline prediction	Measurement	Decision if rejected
Variance grows linearly with horizon	Aggregated variance by non-overlapping block	Replace square-root scaling in limits
Standardized increments are Gaussian	Quantiles and exceedance counts	Add empirical tails or stress scenarios
Increments are independent	Signed and magnitude lag diagnostics	Model state or conditional variance
Parameters are stationary	Rolling estimates with uncertainty	Shorten validity window and recalibrate
Paths are continuous	Gap count after data cleaning	Include jump and execution scenarios

Set tolerances before seeing the later window. Sampling variation guarantees imperfect agreement, so “not exact” is not a useful rejection rule. Use confidence intervals, simulation under the fitted baseline, or a decision tolerance: reject when mismatch changes the operational choice. Preserve the original baseline beside every extension so added complexity earns its place.

Then connect the mismatch to a decision. If ordinary measurement noise fits but tail latency does not, retain the baseline for alert smoothing while sizing timeouts from empirical tails. If drift changes, shorten the planning horizon rather than widening a timeless confidence band. The report should end with the range where diffusion remains adequate, not a verdict that it is simply true or false.

Bachelier’s legacy is a disciplined starting point: independent Gaussian increments, stable diffusion, and explicit horizon scaling. Its value grows when it remains open to rejection.

simple diffusion baseline
  -> predicts square-root spread
  -> predicts Gaussian quantiles
  -> predicts no increment memory
  -> predicts continuous paths

observed violation
  -> name failed assumption
  -> add smallest needed mechanism
  -> retest out of sample

This sequence prevents two errors: finding patterns in ordinary noise and forcing extraordinary behavior into a mild model. Brownian motion is neither foolish nor sufficient. It is a clear ruler whose mismatches teach us where richer finance begins.

Historical baselines deserve primary sources because later summaries often blur arithmetic and geometric models. These links establish Bachelier’s theory and the physical and financial diffusion lineage.

• Benoit Mandelbrot and Richard L. Hudson, The Misbehavior of Markets, Basic Books.
• Louis Bachelier, “Théorie de la spéculation”, the original French paper in Annales scientifiques de l’École Normale Supérieure 17, 1900.
• Albert Einstein, “On the Movement of Small Particles Suspended in Stationary Liquids Required by the Molecular-Kinetic Theory of Heat”, 1905.
• Fischer Black and Myron Scholes, “The Pricing of Options and Corporate Liabilities”, Journal of Political Economy 81(3), 1973.

Brownian motion is useful because it makes strong, testable claims. It becomes hazardous only when those claims remain hidden or are treated as universal.

• Independent increments make variance additive across time.
• Standard deviation grows with the square root of horizon under stable diffusion.
• Drift moves the center; volatility widens the distribution.
• Arithmetic Brownian motion can cross zero; geometric Brownian motion preserves positivity.
• Tails, clusters, jumps, and regime change each reject a different assumption.
• A baseline should organize tests, not overrule observations.

Mastery means using diffusion as a ruler and knowing when it bends. Complete each item without relying on a plotted path alone.

• [ ] I can define a Brownian increment in plain language.
• [ ] I can calculate multi-period standard deviation with the square-root rule.
• [ ] I can separate drift from diffusion.
• [ ] I can distinguish arithmetic from geometric Brownian motion.
• [ ] I can operate both labs and interpret the plotted-curve and surface readouts.
• [ ] I can list the independent, stationary, Gaussian, and continuous assumptions.
• [ ] I can name a domain observation that rejects each assumption.
• [ ] I can explain why a deterministic teaching simulation is not a forecast.