We trained a model on a single year of the rupee, let it predict every year to 2026 blind, then raced fifteen different models against each other to see which forecasting method actually deserves to be trusted.
Three stages, each with the rules fixed before we looked at the answers.
Our source file of RBI daily reference rates was complete except for 2019, 2020 and 2021. We did not skip them. We predicted them like any other year, then checked against independently referenced published annual averages, marked ref throughout, against file for data present in the source. Years 2022 to 2026 are confirmed directly from the file.
Trained on 1999 alone: drift 2.57% per year, volatility 1.56% per year. Then released to predict blind.
The direction was right; the journey was wrong. The model assumed steady 2.6% annual depreciation. Across 27 years that average was almost exactly correct, which is why R² is high. But reality was lumpy: the rupee strengthened to 41 in 2007, crashed in 2008, lurched in 2013. A single calm year could not have known those were coming, and no honest model claims otherwise.
| Year | Simulated | Actual | Error | 90% band | Hit |
|---|---|---|---|---|---|
| 2000 | 44.1 | 44.9 file | -1.8% | [43, 45] | out |
| 2001 | 45.3 | 47.2 file | -4.0% | [44, 47] | out |
| 2002 | 46.4 | 48.6 file | -4.4% | [45, 48] | out |
| 2003 | 47.7 | 46.6 file | +2.4% | [45, 50] | in |
| 2004 | 48.9 | 45.3 file | +7.9% | [46, 52] | out |
| 2005 | 50.2 | 44.1 file | +13.8% | [47, 53] | out |
| 2006 | 51.5 | 45.3 file | +13.6% | [48, 55] | out |
| 2007 | 52.8 | 41.3 file | +27.9% | [49, 57] | out |
| 2008 | 54.2 | 43.4 file | +24.8% | [50, 58] | out |
| 2009 | 55.6 | 48.4 file | +15.0% | [51, 60] | out |
| 2010 | 57.0 | 45.7 file | +24.7% | [52, 62] | out |
| 2011 | 58.5 | 46.7 file | +25.4% | [54, 64] | out |
| 2012 | 60.0 | 53.5 file | +12.2% | [55, 66] | out |
| 2013 | 61.6 | 58.6 file | +5.1% | [56, 68] | in |
| 2014 | 63.2 | 61.0 file | +3.6% | [57, 70] | in |
| 2015 | 64.8 | 64.2 file | +1.1% | [59, 72] | in |
| 2016 | 66.5 | 67.2 file | -1.0% | [60, 74] | in |
| 2017 | 68.2 | 65.1 file | +4.8% | [61, 76] | in |
| 2018 | 70.0 | 65.9 file | +6.3% | [63, 78] | in |
| 2019 | 71.9 | 70.4 ref | +2.1% | [64, 81] | in |
| 2020 | 73.7 | 74.1 ref | -0.5% | [66, 83] | in |
| 2021 | 75.7 | 73.9 ref | +2.4% | [67, 85] | in |
| 2022 | 77.6 | 79.9 file | -2.9% | [69, 88] | in |
| 2023 | 79.6 | 82.6 file | -3.6% | [70, 90] | in |
| 2024 | 81.7 | 83.7 file | -2.4% | [72, 93] | in |
| 2025 | 83.8 | 87.2 file | -3.9% | [74, 95] | in |
| 2026 | 85.5 | 93.3 file | -8.4% | [75, 98] | in |
Its point forecasts were decent. Its confidence was not.
1999 was an unusually calm year, so the model estimated volatility at just 1.56%. True long-run volatility turned out to be 6.40%, more than four times higher. Because it underestimated its own uncertainty, the 90% band was far too narrow and actual values fell outside it roughly half the time.
One year is enough to estimate a trend. It is nowhere near enough to estimate risk. A model that is confidently wrong is more dangerous than one that is honestly uncertain. Measuring how much you do not know is harder, and matters more, than the forecast itself.
A fairer setup: train on everything to 2018 (drift 2.54%, volatility 6.41%, now honest), predict the missing years and confirm the recent ones.
| Year | Simulated | Actual | Error | 90% band | Hit |
|---|---|---|---|---|---|
| 2019 | 69.9 | 70.4 ref | -0.7% | [65, 75] | in |
| 2020 | 71.6 | 74.1 ref | -3.5% | [63, 82] | in |
| 2021 | 73.3 | 73.9 ref | -0.9% | [62, 87] | in |
| 2022 | 75.0 | 79.9 file | -6.2% | [62, 91] | in |
| 2023 | 76.8 | 82.6 file | -7.1% | [61, 96] | in |
| 2024 | 78.5 | 83.7 file | -6.2% | [61, 101] | in |
| 2025 | 80.3 | 87.2 file | -7.8% | [62, 105] | in |
| 2026 | 81.7 | 93.3 file | -12.4% | [62, 109] | in |
Every model trained on 1998 to 2018 only, then predicted 2019 to 2026 with no peeking. Ranked purely on out-of-sample error, never on how well they fit the past.
| Rank | Model | MAPE | RMSE | MAE |
|---|---|---|---|---|
| 1 | GBM (drift) view | 8.53% | 7.80 | 7.10 |
| 2 | Quadratic view | 8.67% | 7.89 | 7.18 |
| 3 | Exponential view | 10.29% | 9.09 | 8.50 |
| 4 | Logistic (S-curve) view | 10.29% | 9.09 | 8.50 |
| 5 | Ornstein-Uhlenbeck view | 10.86% | 10.03 | 9.06 |
| 6 | Random walk + drift view | 11.38% | 10.52 | 9.49 |
| 7 | ARIMA monthly->yearly view | 13.48% | 13.46 | 11.41 |
| 8 | ARIMA(2,1,2) view | 14.13% | 13.42 | 11.86 |
| 9 | Holt (exp smoothing) view | 14.33% | 13.28 | 11.97 |
| 10 | Holt (damped) view | 14.71% | 13.64 | 12.29 |
| 11 | Linear view | 15.22% | 13.35 | 12.57 |
| 12 | ARIMA(1,1,1) view | 16.49% | 15.58 | 13.82 |
| 13 | Naive (last value) view | 17.67% | 16.40 | 14.76 |
| 14 | Gradient Boosting view | 17.68% | 16.40 | 14.77 |
| 15 | Random Forest view | 17.84% | 16.50 | 14.89 |
Click any model for its full explanation: how it works, where it fits, and where it fails.
This is the point most easily missed. Every model on that list is the correct choice for some problem and the wrong choice for others. Random Forest finished last here and would comfortably win on a large tabular fraud dataset. The logistic S-curve failed on a currency and is the right tool for technology adoption. Ornstein-Uhlenbeck was wrong for a trending rupee and is standard for interest rates. Machine learning did not lose because it is bad; it lost because 21 training points and a target moving beyond every value it had seen is the precise condition under which tree models cannot function.
Choosing a model is choosing a shape. The job is matching the shape of the method to the shape of the problem, not reaching for whatever is most sophisticated.
Refit on the full history: drift 3.53% per year, volatility 6.40%, starting from 96.2.
| Year | Median | 90% band | Width |
|---|---|---|---|
| 2027 | 98 | [91, 106] | 15 |
| 2030 | 108 | [89, 132] | 43 |
| 2035 | 128 | [94, 174] | 80 |
| 2040 | 151 | [103, 223] | 120 |
| 2045 | 179 | [113, 281] | 168 |
| 2051 | 219 | [129, 366] | 237 |
The central line is a reasonable extrapolation of a stable 25-year trend, so roughly 219 by 2051 is defensible as a central estimate.
But the honest reading is the band, not the line. By 2051 the 90% range spans 129 to 366. A quarter century of unknown crises and policy shifts lives inside that width. Anyone quoting a single confident 2051 number is not being straight with you.
Yes for trend, no for timing, and it taught us more by failing than it would have by quietly succeeding.
Long-run direction captured with R² 0.87. Beat the naive benchmark by more than half. Survived a blind race against fourteen alternatives and won it.
Missed the timing of every major shock. Starved of data it was badly overconfident, with real values outside its 90% band half the time.