Rail vertical · investor preview

Forecast railway
transformations with math, not optimism.

A purpose-built decision tool for railway holding CEOs and CFOs. Simulate dozens of granular transformation initiatives against 18 real global rail peer cases, with cash-flow-consistent NPV, ROI and payback under realistic shock scenarios. Built first for KTZ-class operators.

Run the rail demo → View on GitHub
18
Global rail peer cases (with promised vs actual)
40
Granular KTZ-specific initiatives
6
Filter dimensions before simulation
1k
Monte Carlo iterations per initiative
From holding-level data to a CFO-ready PDF

How it works

Six stages. Each filters scenarios more tightly. Final output ranks the initiatives that survive on risk-adjusted NPV — with the real cases that backed every estimate.

1

Load the holding twin

Revenue mix by segment, competitor pressure, prior transformations, current strategic priorities. Failed past initiatives flag the same scenarios as elevated-risk.

2

Filter the scenario universe

Across 6 dimensions: strategic objective, investment envelope, time horizon, risk appetite, business segment, implementation model.

3

Match rail peer cases

For each surviving scenario, pull the top 5 most relevant real transformations — only from rail operators globally. No cross-vertical noise.

4

Build empirical priors

Distribution of actual uplift and cost reduction from matched peer cases. Failed projects pull the left tail.

5

Monte Carlo, then ranking

1,000 cash-flow iterations per scenario at 12% WACC. All metrics derive from the same cash flows so the math reconciles.

6

What-if & executive PDF

Live re-run on changed assumptions. One-click executive PDF ready for the board meeting.

Math, not narrative

Methodology

Every number has a derivation. NPV, ROI and payback all come from the same Monte Carlo cash flows — they reconcile.

$ Cash-flow consistent NPV / ROI / payback

One simulation produces cash flows per iteration. NPV is discounted sum. ROI is NPV/investment. Payback is first year cumulative CF turns positive. All three numbers reconcile.

NPV = -Capex + Σy=1..5 ( benefit_y · adoption_y ) / (1 + r)y

Adoption ramps year by year:

0 50% 100% 40%Y1 70%Y2 95%Y3 100%Y4 100%Y5

! Risk modelled, not removed

Each iteration draws from a Gaussian scaled by risk level, then applies independent shocks:

  • Execution failure (from rail peer empirics)5 — 35%
  • Market / commodity downturn8%
  • Cross-border / geopolitical shock12%
  • Cost overrun+10 — 50%
  • Implementation delay+20 — 80%

Per-scenario failure probability is the observed failure rate of similar real rail transformations.

★ Empirical priors from rail peers only

Instead of guessing, TimeStone retrieves what DB Cargo, Trenitalia, Brightline, Indian Railways actually delivered — and what failed at SNCF Fret Cabotage.

prior_mean = w · empirical_mean + (1 − w) · template_mid
where w = min(n_rail_peers / 5, 1.0)

At ≥5 matched rail peers, the empirical mean is used in full. Below that, it shrinks toward a conservative template.

∞ The moat: every engagement compounds

When a recommendation is implemented, the real outcome 12-24 months later is recorded as an immutable OutcomeRecord. The system shifts future priors automatically.

residual = actual − predicted
shift = residual · min(n / 5, 1.0)
prior_mean_next = prior_mean + shift

More real engagements → tighter calibration → impossible for a zero-history competitor to replicate.

The actual decision tool · 2 verticals live

Live demo

Switch between Rail (KTZ) and Banking (Kaspi) verticals — same engine, different peer libraries. 60 seconds for a CEO.

▮▮
Kazakhstan Temir Zholy (KTZ)
$5.5B revenue 130,000 employees 5 segments 4 prior transformations on record
⚠ 2020 digital booking pilot FAILED — similar scenarios flagged

Filter the scenario universe

Choose your strategic constraints. Counter shows how many initiatives survive.
All
Cost reduction
Revenue growth
Service quality
Decarbonization
Resilience
All
< $5M
$5 — 25M
$25 — 100M
> $100M
All
Pilot 6-12mo
Short 1-3y
Long 3-7y
All
Conservative
Balanced
Aggressive
All
Freight
Passenger
Operations
Strategic
Tech foundation
All
In-house
Vendor-led
JV
M&A
Scenarios matching your filters: 50 of 50 in the active library
VENDOR MODE
Vendor optimism stress test. Toggle this on to feed deliberately inflated inputs.
1
Twin loaded
2
Rail peers matched
3
Empirical priors built
4
Monte Carlo running
5
Top initiatives ranked
Ready. Set filters and click Run simulation.

Portfolio synergy · scenario interaction

Initiatives are not additive. When top recommendations share entities, suppliers, customer segments, or talent, their combined NPV is reduced by an empirical overlap penalty.
Sum of individual NPVs
—
Overlap penalty
—
Realistic portfolio NPV
—
Sensitivity analysis — top initiative

What-if · re-run instantly

Stress-test assumptions live. The engine re-runs Monte Carlo in <1s on every change.
12%
1.0x
Standard
Each Monte Carlo run produces NPV, ROI and payback from the same cash flows.
v0.3 preview The simulation engine, next

Entity graph — KTZ as a living system

The next product step decomposes KTZ into ~50 atomic entities. External factors become injectable variables. A transformation propagates through the graph and we evaluate it against 100 synthetic alternate KTZ states.

External factors Depots Fleet Customers Stakeholders
External Depot / Infra Fleet / Wagon Customer Stakeholder Affected by scenario

External factor injection

$80
510
40%
0%
Normal
0%

Node detail

Hover or click any entity in the graph to see its properties, sensitivity to external factors, and which scenarios touch it.
Q3 2026 roadmap — full agent-based simulation with per-entity memory, behavioural logic, and dynamic edges.
v0.4 preview Path-dependent forecasting

Watch the future unfold — 60 months, 100 parallel KTZ timelines

Static averages hide path-dependence. TimeStone runs 100 alternate KTZ timelines forward month-by-month. Random events fire along the way. Each timeline reacts differently.

Cumulative NPV per timeline ($M)
Month 0 / 60

Live cohort statistics

Mean NPV
-
P(NPV > 0)
-
Best timeline
-
Worst timeline
-
Spread (P90-P10)
-
Failed timelines
-
Events fired
0
Current month
0

Event log

Why this matters — A CFO doesn't decide on a single expected NPV. They decide on a portfolio of possible futures. The engine generates 6,000 month-events per run.
METHODOLOGY PROOF Validated on historical cases

Backtest: how TimeStone would have called real transformations

Take a public case. Hide the outcome. Run TimeStone on the pre-project state. Compare the predicted distribution to what actually happened.

Why this is the strongest possible sales proof — A CFO asks 'does your model work?'. We hid outcomes of three real transformations: Hertz ($32M failed), GE Predix ($5B failed), Maersk Spot ($300M succeeded). TimeStone predicted all three within its P10-P90 band. No special tuning.
VALIDATION · 20-CASE BENCHMARK

Out-of-sample validation on 20 real transformations

Every case was scored using only information available before the decision date. Distributions are compared to actual outcomes.

Calibration plot

Predicted vs actual success rate per bucket. Closer to diagonal = better calibrated.

Coverage check · per case

Each row: P10–P90 band + actual NPV marker (cyan inside, red outside).

All 20 validation cases

Pre-project state, TimeStone counterfactual forecast, actual realized outcome, hit / miss verdict.

Brier score · vs naive baselines

Brier = mean((predicted − actual)²). Lower is better.

Leave-one-out cross-validation

For each of the 20 cases we re-fit priors on the other 19 and predict that case from scratch. R² measured on the LOO sample.

P10–P90 uncertainty: TimeStone vs McKinsey published ranges

Industry consultancies publish point estimates with wide ±50% bands. TimeStone returns a quantified P10–P90. Lower ratio = tighter, more decision-useful band.
McKinsey / BCG range TimeStone P10–P90
PUBLIC COMMITMENT · IMMUTABLE

Pre-registered predictions on 8 ongoing transformations

Backtests on the past are convincing. Forecasts on the future are accountability. Below: TimeStone commitments on 8 ongoing transformations, committed to git with timestamp.

⏱
Immutable since 20 May 2026
All 8 predictions below are stored in
predictions/2026-05-20.json — append-only. No edit history allowed. Outcome verification will be appended as a separate file in 2027–2029.
The rail corpus

Rail peer library

18 real railway transformations from operators globally. Cross-vertical cases are deliberately excluded — they were noise, not signal.

The flywheel

Why a generic LLM cannot replicate this

Three structural advantages stack and compound over time.

  • Vertical-specific case libraries — rail peers for rail clients, banking peers for banking clients. No cross-vertical pollution.
  • Append-only OutcomeRecord database. Every real engagement adds a 'predicted vs actual' pair.
  • Failure-mode aware retrieval. If KTZ already failed at digital booking in 2020, the model flags similar scenarios as elevated risk.
  • Region-specific peers cannot be extracted by a public LLM at this quality.
  • Open-source engine, closed-source data. Code is auditable; the moat is the curated proprietary corpus.
FAILEDSNCF Fret Cabotage · 2018
€340M digital freight platform — derailed by union blockades and B2B sales adoption gap.
SUCCESSBHP Mt. Newman · 2010-2020
−13% maintenance cost, +18% throughput on iron-ore corridor after $1.5B autonomous rail rollout.
PARTIALDB Cargo myRailLog · 2017-2022
Technology delivered. Adoption stalled at 30% of corporate shippers — relationship sales still wins.
▲ These three signal patterns directly shape KTZ recommendations.
ABOUT

Who's behind TimeStone

AT

Arman Torebek

Founder · CEO
Built TimeStone after years inside corporate transformation consulting in Kazakhstan and the CIS — seeing first-hand how the same patterns of optimistic vendor pitch decks, hidden execution risk, and post-mortem write-downs repeated across every industry. Background in finance, consulting and product. Operates from Almaty.
GitHub Email

Why TimeStone exists

Every large transformation goes through the same ritual. A vendor presents a deck. The board approves nine figures. A year later someone discloses an impairment. McKinsey publishes another paper saying 70% of transformations fail. The cycle restarts.

TimeStone was built on one belief: this loop persists because there is no decision-time instrument that uses real outcome data from comparable transformations. We have the data — it's in 10-K disclosures, lawsuit filings, HBR post-mortems, IPO prospectuses. What's missing is a rigorous engine that consumes it and produces a calibrated forecast.

"Most forecasting tools in this space are intellectual luxuries. We built one a CFO can actually defend to their board."

The product is open-source on GitHub. The case library is curated in plain JSON. The math is documented and unit-tested. The methodology is reproducible. Anyone with a Jupyter notebook can verify our calibration metrics. Investors and clients receive open-book trust by design.

Advisor seats open

Industry advisorEx-CFO of large CIS industrial holding
Quantitative advisorPhD finance / decision science · validates methodology
Vertical lead · bankingSenior digital banking practitioner — Q3 2026
Open source under the hood

Install & run on your data

The Python engine generating these forecasts is open source. Clone, install, run on your data. 

# Clone & install
git clone https://github.com/westfellow25/timestone-ai
cd timestone-ai
pip install -e .

# Run a full assessment (produces PDF + JSON artifacts)
python -m timestone assess "Kazakhstan Temir Zholy (KTZ)"

# Process recorded outcomes and recompute calibration
python -m timestone calibrate

# Or launch the interactive dashboard
streamlit run src/timestone/interfaces/web/dashboard.py