261 out-of-sample forecasts of the rand, one every Friday from 1 January 2021 to 31 December 2025, at a 28-day horizon, scored against an AR(1) benchmark.
Built for the SES x Codera Analytics Hackathon in July 2026 and extended afterwards. Throughout this repository "we" refers to the competition entry (myself and Stefan von der Heyden) and "I" to the statistical work, test suite and documentation added after it.
| Forecast combination (our model) | RMSE 0.5600 |
| AR(1) benchmark | RMSE 0.5624 |
| Improvement | 0.42 per cent, with 4 of 5 models ahead of the benchmark |
| Target series | SARB EXCX135, checked against the SARB Web API across 9 141 observations, maximum difference 0.000000 |
| Reproducibility | data/processed deleted and rebuilt from raw, maximum RMSE difference 2.22e-16 |
| Five minutes | The next two sections |
| Twenty minutes | notebooks/walkthrough.ipynb, the analysis end to end with charts, already executed |
| An hour | docs/METHODOLOGY.md, the full write-up |
| If you doubt the numbers | data/processed/forecast_log.csv, all 261 forecasts, every model, recompute the headline in one line |
| If you doubt the discipline | pytest -q, fourteen tests that fail if any forecast could not have been made on the day it claims |
At a one-month horizon the rand is close to a random walk. Since Meese and Rogoff (1983) structural exchange-rate models have lost to a no-change forecast with some regularity, and honest published gains sit around 1 to 5 per cent in RMSE. We did not try to fight that. The model is a combination: a random-walk anchor, a carry tilt from the JIBAR minus US T-bill gap, and a heavily shrunk ridge on commodity, dollar and risk factors, weighted by how accurate each one has been recently.
It beats the AR(1) by 0.42 per cent. The more interesting part is that we can say why.
An AR(1) fitted to this series estimates persistence between 0.9971 and 0.9986.
Close to one, but below it, and anything below one pulls every forecast toward an
implied long-run mean of c / (1 - phi). So the question is where that mean sat
relative to the market.
We refitted the AR(1) at all 261 origins and recomputed it each time. The implied mean sat below the prevailing spot rate at every one of the 261 origins, by R2.11 on average and never by less than R0.57, while the rand was climbing from R14.62 to a peak of R19.74.
A mean-reverting model applied to a trending series pulls every forecast the wrong way. RMSE squares errors, so the occasional large miss that produces does most of the damage to the score.
That is the whole finding. Our models do not predict the rand better in any deep sense. They decline to make one specific mistake, repeatedly. Which also tells you when the advantage should disappear, and it does: 2025 goes to the benchmark, because 2025 is the year the rand moved back toward the mean.
The improvement is not statistically significant, with a Diebold-Mariano p-value of 0.47. Left there, that is a weakness. The more useful question is what significance would have taken.
| Origins available | 261 |
| Observed improvement | 0.42 per cent |
| Minimum detectable at 5 per cent with 80 per cent power | 1.64 per cent |
| Origins needed to detect an effect this size | 3 906, roughly 75 years of weekly forecasts |
A non-significant result on 261 origins is what you should expect at this effect size, rather than evidence that the model does nothing. It also means a competing entry reporting significance on a sub-1-per-cent improvement over the same window is claiming something its data cannot support.
We also ran gradient boosting under identical rules as a control, expecting it to lose, and it lost by 7.5 per cent. In the full pairwise DM matrix that is the only difference anywhere that reaches significance. Keeping a failed model in the results is more informative than dropping it.
The fourteen tests re-derive the leakage guarantees from the raw files rather than trusting the pipeline that produced the processed ones. And section 12 of the methodology lists four errors the audit caught, including a bug in my own statistics module that understated a headline figure by twelve percentage points.
The part worth checking if you are evaluating this.
- Information sets. Every predictor enters through a backward
merge_asof, so a forecast made on Fridaytsees only observations dated on or beforet. - No publication lags. Every predictor is a daily market price or rate, never revised and never delayed, which makes look-ahead bias impossible by construction rather than merely avoided. We gave up CPI and the current account to buy that.
- Resolved-only training. The regressions train only on origins whose 28-day
outcome had already resolved:
tr = tr[tr.index <= i - H]. - Embargoed weights. Combination weights use only errors from forecasts resolved 28 days before the current origin.
- Frozen configuration. Feature set, intercept choice, decay and sharpness were
chosen on a 2015-2020 validation era whose targets all resolve before 2021,
then frozen. The whole search grid is committed in
validation_metrics.json, so you can see how many configurations were tried and not only which one won.
| Model | RMSE | vs AR(1) | What it is |
|---|---|---|---|
| AR(1) | 0.5624 | — | The required benchmark. OLS with intercept on the weekly log level, expanding window from 1999, iterated 4 steps. Scored against the better of the log and level variants. |
| Random walk | 0.5600 | +0.43% | No change. The strongest single benchmark at this horizon. |
| Combination | 0.5600 | +0.42% | Our submission. Inverse discounted-MSE weighted average of {RW, carry, ridge}, decay 0.97, sharpness 2. |
| Ridge | 0.5605 | +0.34% | 4-week log return on the compact predictor set. Standardised in-window, penalty by time-ordered CV, no intercept, so heavy shrinkage collapses it into the random walk instead of a stale mean. |
| Carry | 0.5606 | +0.31% | Covered-interest-parity forward rate from the JIBAR minus US T-bill gap. |
| Gradient boosting | 0.6046 | −7.51% | Depth-2 comparator under identical discipline, kept as a control. |
├── README.md
├── notebooks/
│ └── walkthrough.ipynb the analysis end to end, executed
├── docs/
│ ├── METHODOLOGY.md the full write-up
│ └── original_task_brief.pdf the competition brief
├── src/
│ ├── fetch_data.py 1. download every raw series
│ ├── verify_provenance.py 2. EconData against the SARB Web API
│ ├── build_dataset.py 3. weekly Friday dataset, backward as-of merges
│ ├── run_backtest.py 4. walk-forward out-of-sample backtest
│ ├── statistical_tests.py 5. DM matrix, power analysis, rolling accuracy
│ ├── make_figures.py 6. report figures
│ ├── make_presentation_figures.py
│ ├── validate_results.py 8. audit every plotted number against raw data
│ ├── make_report.py 9. build the technical report PDF
│ ├── build_notebook.py regenerate the walkthrough notebook
│ ├── make_presentation.js rebuild the slide deck (Node)
│ └── run_all.py the whole pipeline in order
├── tests/
│ └── test_no_lookahead.py fourteen leakage and integrity tests
├── data/
│ ├── raw/ exactly as downloaded, never hand-edited
│ └── processed/ weekly.csv, forecast_log.csv, metrics.json,
│ statistical_tests.json, validation_metrics.json
├── figures/
└── reports/
├── technical_report.pdf
└── presentation.pptx
git clone <this-repo> && cd <this-repo>
pip install -r requirements.txt
python src/run_all.py # data, backtest, statistics, figures, report (~5 min)
pytest -qNo API keys and no paid data. The raw files are committed, so everything from
build_dataset.py onwards runs offline. Rerunning reproduces every published
number to twelve decimal places.
Rebuilding the slide deck also needs Node: cd src && npm install && node make_presentation.js.
| Series | Source | Role |
|---|---|---|
| USDZAR (EXCX135) | SARB via EconData, MARKET_RATES 1.0.0, EXCX135.B.A |
Target |
| JIBAR 3M | SARB Web API | Carry |
| US 3M Treasury bill | FRED (DTB3) |
Carry |
| Gold, platinum, Brent | SARB, NYMEX, FRED | Commodity block |
| Broad US dollar index, VIX | FRED (DTWEXBGS, VIXCLS) |
Dollar strength, global risk |
SARB. MARKET_RATES. Version 1.0.0. Pretoria: SARB [producer]. Johannesburg: EconData [distributor], 2026.
Statistical significance, for the reasons above. Directional predictability, which sits near 50 per cent for every model, so nothing here calls which way the rand moves. Profitability, which was never tested, and which is a different objective from RMSE. Persistence of the advantage, since 2025 already goes the other way.
The direction I would take next is option-implied volatility and risk reversals, following Greenwood-Nimmo, Steenkamp and van Jaarsveld (SARB WP/22/14), and then density rather than point forecasts. For anyone actually hedging currency risk the distribution is the decision-relevant object, and forecasting dispersion in a series whose mean is nearly unforecastable is the better-posed problem.
Aakash Anil, BCom Honours Economics, Stellenbosch University, with Stefan von der Heyden, entered as team Spurious Decision. The entry did not place. I rebuilt the repository afterwards as a portfolio piece and added the statistical testing, power analysis, tests and documentation at that point.
Anthropic's Claude was used as a coding assistant for pipeline code, literature search and typesetting. The modelling decisions, validation design and final specification are mine. OLS, ridge and weighted averages were chosen partly so that every component can be derived by hand.
MIT licensed. See LICENSE for the data terms, which differ.